OpenRouter now serves the Codex and Claude Code model lists, so
ori codex, ori claude and their setups stop building or narrowing a list on the laptop. The desktop app gains chats on the local Ori Code agent, native routines and connections, and updates that install themselves. Slack replies hold together through long turns, lost responses and deleted messages. Backups become labeled versions that you list and restore by name. The built-in harness defaults to openrouter/auto, keeps session history for 30 days, and refuses more risky local inputs.Breaking changes
- Cross-family model settings retired.
ORI_CLAUDE_CROSS_FAMILY_MODELSandORI_CODEX_CROSS_FAMILY_MODELSno longer have an effect from any source.ori claudeandori codexkeep other families’ models, and a launch no longer refuses a model for its family.ori doctorwarns about a key that is still set locally. - Codex 0.164.0 or later.
ori codexandori harness setup codexrefuse an older Codex withORI_CLI_HARNESS_VERSION_UNSUPPORTEDand an upgrade hint.
Codex and Claude Code
- Model lists from OpenRouter. OpenRouter applies a Managed Machines
ORI_CLAUDE_MODELSorORI_CODEX_MODELSlist for every launch, and the pickers show only the usable listed models in list order. OpenRouter refuses an unlisted model. A list set in managed preferences or/etc/ori/config.jsonstill narrows on the laptop. - Codex reads the served list. Codex fetches its picker from OpenRouter, so Ori writes no local Codex model catalog. The next start through setup removes a catalog line that an older setup wrote, and prints one line saying so. Long Codex sessions now compact before they reach the model’s input limit.
- Claude Code default. With
ORI_CLAUDE_DEFAULT_MODELset, the Default row, the model that setup saves and the modelori claudestarts on are the first model OpenRouter serves. An empty served list leaves only the default row. - Saved Codex logins.
ori harness setup codexnow hides every saved login in~/.codex/auth.json, asori codexalready did, so the Codex app can generate images. Setup andori doctor codexsay how to use the ChatGPT plan in plain Codex. - Claude Desktop import. Claude Desktop in gateway mode turns on the claude.ai import.
- Vercel fx. The new
ori fxlauncher starts Vercel’s fx coding agent on OpenRouter.
Desktop app
- Chats on the local agent. A desktop chat can run on the local Ori Code agent in a folder you pick. A models mode chats with any model through the same agent.
- Turn view. Turns show a one-line live status, narration, diffs, a files panel that tracks changed files across repositories, and branch commits. The “Worked for” time stops when a turn ends.
- Routines. The desktop app manages routines natively, including one-off routines, routines made by describing them, and routine cards in chat. A missed one-off says that it did not run, and you edit it from its due time.
- Interns. The desktop app manages an intern’s features, secrets, sharing, connections, library and change history. It asks before it suspends an intern, and it keeps an intern whose delete failed visible with a retry.
- Dashboard and accounts. The desktop app connects the dashboard with Ori’s own sign-in in the default browser, and mirrors the dashboard’s connections and your accounts.
- Library and Artifacts. The Library is now called Artifacts. An artifact opens in a new chat, and HTML files from chat render with a code toggle.
- Composer and replies. You can queue and steer messages, quote and comment on parts of a reply, and use Markdown, LaTeX and list continuation in your own messages. Images and files that an intern makes open from its reply.
- Updates. The signed macOS app updates itself from the alpha channel. It asks before it downloads an update, shows progress, and offers Install and restart once the update is ready.
- Debug info. Help > Export debug info and Settings > About write a debug report zip to Downloads and show it in Finder. The report holds no keys.
- Fixes. Header buttons and window dragging work again, Slack mentions show as names, and markdown code blocks no longer mark every line.
Slack
- Long replies stay in one message. A long turn keeps its reply in one message past Slack’s 10-minute stream limit, past 50 plan cards and after a failed stream append.
- Lost and deleted messages. A message posts once when its response is lost, and a reply stream ends quietly when someone deletes its message. A stream call that never connected is sent again.
- Steering and stopping. Quick messages in a row leave one steer pickup note. Stopping a thread also drops the messages that wait behind the running turn.
- Threads. A model override applies only to the thread it was set in. Ori no longer renames a thread that the intern started, and a muted thread stays muted when its state read fails.
- Formatting and size. Messages started in Slack keep their formatting, long answers break on a whole character, and the 100 MiB per-turn attachment limit is gone.
Hosted interns and schedules
- Feature commands everywhere. Slack runs, schedule runs, run now and live-edited fires can use feature commands.
- Routines. Interns get a routine tool, the time of the turn, and one routines skill. A missed one-off is recorded instead of re-armed for the next year.
- Skills. A skills tab and an add skill action manage skills, and slash commands run skills, also when typed mid-sentence.
- Reloads. A busy intern applies feature edits after the runs already in flight finish.
- Turns survive interruptions. The built-in harness records each turn, request and tool call, and resumes an interrupted turn from that record.
Backups and versions
- Labeled versions.
ori backup --label "<description>"saves a backup as the next version, such asv1orv2, and pushes the label whenoriginexists.ori backup listshows each version with its parent, time, creator and description. - Restore by version.
ori restore v2restores that version and commits it forward.ori restorerefuses a name that is also a branch or tag withORI_CLI_RESTORE_REF_AMBIGUOUS. - Versions on the daemon. The daemon lists, labels and restores versions over its API. A hosted intern knows its versions and points the person to the Backup section of its settings.
- Build reviews. When the owner asks the intern to review how it is built, the intern saves a version first. It then rebuilds itself from a list of what it can do and replays recent real requests. “Go back to version v4” undoes the review.
CLI and built-in harness
- Default model. The built-in Ori harness uses
openrouter/autowhen nothing names a model.ORI_MODELstill overrides it. - Session history. On a person’s machine, a session stays resumable for 30 days after its last turn instead of 72 hours.
ORI_SESSION_HISTORY_RETENTION_DAYSchanges the window. Deployed interns keep 72 hours. - Credit and key errors. The built-in harness says when you are out of credits and why OpenRouter refused. It names key-limit
402errors, also mid-stream, and links to the key’s page. - Windows shell. Ori finds Git Bash on Windows and reports it in
ori doctor. - Chat TUI.
/approvalsis user-only, so the model cannot switch itself to self-drive. The session picker shows better names.
Security
- Project dotenv. Ori refuses project dotenv variables that change what child processes load.
- Repository MCP servers. A repository HTTP MCP server that sends headers needs trust first.
- Protected paths. The built-in harness refuses native writes and edits inside
.ori, and isolates Linux shell commands from host Unix sockets. - Dev host.
ori devrefuses a discovered.ori/dev.jsonthat names a non-loopback host.
The desktop app gets a run settings menu in the composer, connect cards in replies, an Interns table, and spaces that follow you between desktop and web. Slack keeps a chat model switch for later turns, and hosted interns finish scheduled tasks after a restart. At
ORI_TELEMETRY_LEVEL=off, Codex keeps its local plugins, MCP plugins and the Codex app tools. On Windows, ori harness setup codex no longer refuses when PowerShell starts slowly. 0.16.2 was not published, so this release also carries every change listed in 0.16.2.Desktop app and web
- Run settings in the composer. A run settings button in the desktop composer opens the model, reasoning, and routing choices for the current chat.
- Connect cards. A setup link in an intern reply shows as a connect card in the desktop chat.
- Interns table. The desktop app shows interns in a table view with a toolbar and pages.
- Account row. The desktop sidebar shows an account row at the bottom.
- Spaces on every surface. Spaces follow you between the desktop app and the web. The desktop window keeps its position and size.
- Connection health. The desktop Connections view says when a GitHub connection has no current access token, so requests fail until the token refreshes.
- Chat layout. Desktop and web chats use a narrower column with larger, full-width replies. The loading placeholder matches the loaded chat.
Hosted interns
- Scheduled tasks after a restart. A scheduled intern task that a restart or a dropped connection interrupts now finishes its work. A task that a person cancels stays cancelled.
- Linked Google accounts. Before a Google task, a hosted intern checks which Google accounts the member linked. When the member’s email is missing, the intern asks the member to link Google from the dashboard.
- Refused MCP servers. When an intern refuses a stdio MCP server, the reply says the server “is not available on this machine yet” and links to its dashboard connection. Local
oriwithout a dashboard keeps its message. - Announced work finishes. On models without commentary phases, an intern finishes the work it announces instead of stopping after a “Checking” line.
- More parallel runs. Interns on the built-in Ori harness get 640 active-run slots by default instead of 64. A lower value that an administrator set still applies.
Slack
- Chat model switch sticks. Asking an intern in chat to change its model now changes the model for every later turn in every thread and DM, like
!model.!model resetgoes back to the dashboard model. - Footer names both models. The reply that switches models names the old and the new model in its footer. Later replies name only the new model.
- Spawned thread anchor. When an intern spawns or forks a thread, the anchor reply in the original thread shows only the View spawned thread button.
- Emoji at the cut. A long label that Slack cuts exactly at an emoji now keeps the whole emoji before the ellipsis.
Codex
- Local plugins at telemetry off. At
ORI_TELEMETRY_LEVEL=off,ori codexand Codex afterori harness setup codexkeep local plugins, skills and plugins that run their own MCP server, and the Codex app keeps computer use, the in-app browser and its thread tools. Only the parts that reach OpenAI stay off: remote plugins, plugin sharing, recommended plugins, tool suggestions and apps.ori doctor codexwarns when a setup from an older release still turns local plugins off, and runningORI_TELEMETRY_LEVEL=off ori harness setup codexagain turns them back on. - Windows setup. On Windows,
ori harness setup codex,ori harness reset codex, andori doctor codex --fixcheck whether the Codex app runs withtasklist.exefirst. Setup no longer refuses withORI_CLI_HARNESS_LAUNCH_FAILEDwhenpowershell.exestarts slowly on a busy machine.
Allowed commands now take command paths such as
harness setup, and a command that the saved list refuses downloads the current list before it refuses. ori harness setup codex prints a short summary, and Codex and Pi requests show as their own apps in Logs. ori support-bundle collects a launch’s logs at telemetry off. The desktop app gets an intern switcher on the new-chat greeting and many input and list fixes.Managed machines
- Command paths in Allowed commands. An
ORI_ENABLED_COMMANDSentry can be a command path such asharness setup. It allows that command and every command under it, and nothing else underori harness. An entry that names an argument, such asharness setup codex, allows nothing, and the refusal says that harness names go inORI_ENABLED_HARNESSES. An ori older than this release refuses the whole parent command, so setORI_MINIMUM_VERSIONbefore you use a path entry. - Cursor setup follows the harness list.
ori harness setup cursornow checksORI_ENABLED_HARNESSESlike the other setup harnesses, and the dashboard can listcursoras an allowed harness. - Fresh list before a refusal. With a saved login, a command that the saved list refuses downloads the launch settings once and runs when the new list allows it. Offline, after 1500 ms, or when the new list still refuses it, the command refuses as before.
- Fresh list after an allowed command. An allowed command downloads the launch settings in the background, so the next command uses the new list. The command’s own output is not delayed. Before it exits, ori waits at most 1500 ms for the download.
- Downloads ori skips. Ori does not download the list when there is no saved login, when the list comes from managed preferences or
/etc/ori/config.json, or when a project.envpoints the OpenRouter URL or transport somewhere ori does not trust. An OIDC login never renews its token for this download. - First launch checks the dashboard list. On a machine with no saved launch settings,
ori codexand the other launches refuse withORI_CLI_MANAGED_COMMAND_DISABLEDwhen the downloaded list leaves them out. - Release line pin.
ORI_TARGET_VERSION, an administrator-only setting such as0.16, keeps machines on the newest stable patch of that line.ori update,ori update --check, and auto-update install only that line. An ori on a newer line installs the target before it launches a harness. - Ignored saved keys. Each launch-settings download logs one warning for each saved key that ori drops. The warning names the key and the reason, never the value.
- Codex launch settings text. Each native
codex.*launch setting now says that it applies toori codexonly, not to the Codex app or to plaincodexafter setup.
Codex and Pi
- Short setup output.
ori harness setup codexprints one line for the change, one line for where the Codex CLI and the Codex app now send requests, and the undo command. The other lines print only when they are true, such as a ChatGPT login or a replaced earlier setup. - Previous setup line. The
codex --profile originalline prints only when there was a real previous setup. A machineconfig.tomlthat setsmodel_providernow counts as one. - Removed setup notes. Setup no longer warns about ChatGPT Work chats. The notes about hidden Chat, voice, and Files, and about OpenAI-hosted plugins, moved to
ori harness setup --helpand the--jsonmessages. - App names in Logs. Requests from
ori codex, and from the Codex app and Codex CLI afterori harness setup codex, show as Ori Codex. Requests fromori pishow as Ori pi, and Pi sends its own attribution headers again. A machine that is already set up gets the change when you runori harness setup codexagain. - Codex feedback at telemetry off. At
ORI_TELEMETRY_LEVELoff,ori codexturns off Codex feedback, andori harness setup codexwrites[feedback] enabled = false. An existing setup atoffshows the “predates telemetry off” notice until you run setup again.
Claude
- Sonnet 5.5 and Haiku 5.5 defaults. With no model catalog,
ori claudesets the Sonnet tier and the auto mode model toclaude-sonnet-5-5[1m]and the Haiku tier toclaude-haiku-5-5[1m].ori harness setup claudewrites both models into Claude Desktop. - Claude Desktop slots from the catalog. Each Claude Desktop model slot takes the newest Claude model of its family from the workspace catalog. A family that the catalog leaves out gets no slot. The daily catalog refresh moves the slots, and Claude Desktop shows them after its next full relaunch. An administrator’s
ORI_CLAUDE_<TIER>_MODELstill names its slot. - Unreadable Claude Desktop file. On macOS,
ori harness setup claude-desktopstops withORI_CLI_HARNESS_CONFIG_WRITE_FAILEDand changes nothing when it cannot read a file inconfigLibrary. Before, setup left that file out of the backup.
CLI
- Support bundles. Every
ori <harness>launch writes a small record under~/.ori/logs/launches/, and ori keeps the newest 50.ori support-bundlewrites a local folder with the launch record and, for Codex, the time, level, and component of each Codex log line. It never uploads anything. - Scrubbed bundles.
ori support-bundle scrub <folder> --terms terms.txtreplaces each listed term with<term-N>. The placeholders match across bundles. A terms file that conflicts with an earlier scrub fails withORI_CLI_SUPPORT_BUNDLE_SCRUB_MAP_CONFLICT. - Launch failure hint. A refused launch with no hint of its own now ends with a line that names
ori support-bundle. - Intern credential error.
ori atandori interns attachfail withORI_CLI_INTERN_SECRET_KEY_MISMATCHwhen OpenRouter cannot open the intern’s stored credentials. The message says that a restart does not fix it. - Doctor messages.
ori doctor <harness> --fixfor a harness other thancodexnames both repairs that--fixcan do. The key expiry inori doctorcounts calendar days in your local time zone. - Profile help.
ori profile list --helpsays that human output marks the active profile(active).
MCP servers
- Global file from the home folder. When the home folder is not a git repository,
ori mcp list,ori mcp test,ori code, andori devtreat stdio servers in~/.ori/mcp.jsonas global, also when they run from~. A home folder that is a git repository, and every unattended run, keep the repository rules. - Stdio servers on vault interns. On an intern that uses the vault, a stdio MCP server now reaches the network through the sidecar, like the agent’s shell.
- Approval link for hosted interns. A hosted intern that needs an unapproved stdio MCP connection replies with a link that opens the connection in the dashboard.
ori codeon a laptop is unchanged.
Chat TUI
- Choice prompt countdown. In
ori codeandori tui, each key in a question prompt restarts its 30 second auto-answer. The prompt header shows the countdown, such asInput requested · auto-answers in 24s.
Desktop app
- Intern switcher on the greeting. On a new chat, click the intern name in the greeting to pick another intern and start a chat with it. The list opens on your current intern and works with the arrow keys, Enter, and Escape. Keyboard focus now lands in the sidebar’s Switch intern list.
- Input methods. With an IME on, Enter confirms the word and Escape cancels the conversion. Neither key saves, picks, cancels, or closes anything while the composition is open.
- Hung interns. A hung intern no longer freezes the intern and session lists. Its reads and writes fail with “took too long to answer” after 30 seconds. Turns and steering are unchanged.
- Running turns. A working session keeps following its turn when one session-list read fails. Before, the app showed “turn finished” for a turn that was still running.
- Load errors. The Team view’s “Couldn’t load threads” row says why, one line for each reason. The app asks about a failing intern once instead of every few seconds, and Retry asks again at once.
- Thread library. With Group by Day and any sort, each day header shows once, newest day first.
- Tool step names. MCP tool steps name their server, for example “List issues Linear” instead of “List issues mcp linear app”.
- Intern names. The New intern form removes accent marks from the suggested name, so “Renée” gives
renee.
Slack and hosted interns
- Long answers. An answer longer than 12,000 characters posts cut to fit and ends with a truncation notice. A failed, cancelled, or restarted answer still ends with its status line, also when its error text is very long.
- Subagent output. Subagents that a tool starts keep their output with that tool. Only the main thread’s messages reach the Slack reply.
- Slack restart guidance. A hosted intern whose Slack connection cannot start sends the person to the Slack card and its Restart button, and says when Restart cannot help.
Feature authoring
- Signed button clicks. A handler registered with
use("@ori-builtins/slack").onButtongetsclick.signedRequest, the exact body, signature, and timestamp that Slack signed. It isundefinedwhen the delivery had no signature headers. - Decisions client.
fetchOpenRouterDecisionsfromoriasks OpenRouter Decisions typed questions about a state and returns typed answers with the served model, tokens, and cost. A call inside an operation is billed to that operation. - Percent tolerance. A percent
toleratedFailuresonfanout, such as"29%", allows exactly that share of failed branches. - Package versions. A feature module that imports a package that is only in the Bun store loads the highest installed version.
Observability
- Event loop delay.
ori.event_loop.delayno longer includes the sampler’s own run time, so the constant floor of about 18 ms is gone. The resource gauges have values from daemon start. - Evicted sessions.
ORI_SESSION_HISTORY_NOT_FOUNDlogs at warn level, not error level.
ori harness setup codex now warns about every way the Codex app can bypass OpenRouter, and it no longer stops intermittently when the app is open. ori harness setup claude backs up the Claude Desktop configuration before it changes it. ori alias follows native updates, and ori auth region names each harness to set up again. Security fixes harden webhooks, desktop links, remote features, and the sandbox of ori code --approvals manual.Codex setup and Codex Desktop
- ChatGPT login warning. At
ORI_TELEMETRY_LEVELno-contentorfull,ori harness setup codexwarns when the Codex app is signed in to ChatGPT. Those chats may run on ChatGPT’s own service and bypass OpenRouter. The hint says to sign out in the app, or to runcodex logout. CODEX_ACCESS_TOKENwarning. At any telemetry level, setup warns whenCODEX_ACCESS_TOKENis set. The Codex CLI in that shell then uses a ChatGPT token. The hint says to unsetCODEX_ACCESS_TOKEN.- ChatGPT Work warning. When the Codex app is installed, setup warns that ChatGPT Work chats do not reach OpenRouter. The hint says to pick Codex under Switch mode in the app’s sidebar. The warning also appears in the
--jsonmessages. - Hidden app features. When the Codex app is installed, setup says that the app hides Chat, voice, and Files while it is routed through OpenRouter. After
ori harness reset codex, a ChatGPT sign-in shows them again.ori harness setup codex --helpsays the same. - App restart on macOS. With the Codex app open,
ori harness setup codex,ori harness reset codex, andori doctor codex --fixquit the app, apply the change, and open it again. Before, they sometimes stopped withORI_CLI_HARNESS_ALREADY_RUNNING. Claude Desktop and Cursor setup now also find a running copy outside the usual folders. CODEX_HOMEnote.ori doctor codexprints the “CODEX_HOME is set” note only whenCODEX_HOMEpoints somewhere other than~/.codex.- Stable session ID.
ori codexkeeps the same session ID when a Codex home that a symlink names is created later.
Claude
- Claude Desktop backup. On macOS,
ori harness setup claudeandori harness reset claudecopyclaude_desktop_config.jsonandconfigLibrary/to~/.ori/claude-desktop/backups/<time>before they change anything. The first backup goes tobackups/before-ori, which Ori never overwrites or prunes. Ori keeps the newest backup and the nine before it. - Failed backup. When the backup fails, setup stops before it changes Claude Desktop. Reset warns and still restores Claude Desktop, because reset is the recovery path.
CLI
- Aliases follow native updates. After
claude update, the nextori clauderuns the new version and moves the binary alias to it, so plainclauderuns it too. This works for every launcher that links into aversions/folder, such as cursor-agent. - Stale alias check.
ori doctorwarns when a binary alias runs an older version than the one installed.ori unalias --binaryrestores the launcher at the version the alias last followed. A launch no longer takes a lock, so it never blocksori alias,ori unalias, or setup. - Region change notice.
ori auth region <region>prints one warning for each set-up harness whose address the change moved. Each warning names the setup command to run again, and theCLAUDE_CONFIG_DIRorPI_CODING_AGENT_DIRwhen it differs. - Closed output pipes.
ori --human <command> | head -Nnow exits on its own. Ori drops the lines after the pipe closes and finishes the command, so setup still writes its files.
Security
- Manual approvals sandbox. With
ori code --approvals manual, a sandboxed command has no controlling terminal. On macOS, it cannot resolve names or open URLs through the system services when egress is off. On Linux, it cannot read~/.ori/credentials.json. - Webhooks under load. A mounted webhook keeps verifying real deliveries while up to 128 clients send slow bodies. The next concurrent body read gets
429. - Desktop links. The desktop app opens only
https:links andhttp:links tolocalhost,127.0.0.1, or[::1]. It refuses every other link withurl_refused. - Remote features.
ori features addrefuses a git ref with an empty,., or..segment, a leading-, or a trailing/or.lock. The SSH fallback refuses an unknowngithub.comhost key and says how to check it. - Deploy image Bun.
ori init --with-dockerandori update --sync-dockernow writeFROM oven/bun:1.4.4-slim AS runtime, a Bun with the Brotli fix. - Connection errors. An OIDC renewal that cannot reach the issuer is retryable again. A vault TLS failure reports
EGRESS_SIDECAR_VAULT_TLS_FAILEDagain.
Automations
- Long summaries. A delivered automation summary or chat reply that is longer than the delivery limit now ends at the last full sentence, followed by
_Truncated: the reply exceeded the delivery limit._. - Slow Slack delivery. A delivery that needs Slack retries now records
deliveredin the run history, instead oftimed out.
Harness versions
- Verified versions. This release verifies Claude Code 2.1.309, Claude Desktop 3.6556.2, Codex CLI 0.164.0, Codex Desktop 26.1016.62342, and Pi 1.4.0.
ori codex never changes your Codex files at launch. Run ori harness setup codex once when you want Codex, Codex Desktop, and ori codex to share one OpenRouter provider and one thread list. Administrators can set native Claude Code, Codex, and Pi settings, keep a harness on one model family, set the telemetry level, and require a minimum Ori version. Slack replies stream with tool steps in a plan, ori workflows runs durable feature workflows, and ori eval prompt compares models without an eval file. Every command prints in one output style, sign-in asks for the gateway, ori auth shows the remaining budget, and saved credentials move into one encrypted store.Breaking changes
- Slack assistant pane retired. Ori no longer listens for the events of Slack’s assistant pane, and the prompt no longer names the channel that a pane opened from. The first message of every thread, in a direct message or a channel, now names the thread’s session in at most 200 characters.
ori artifactremoved. Theori artifact save,ori artifact list, andori artifact deletecommands are gone from the CLI. An intern that loads the feature catalog gets the same commands from the catalog’sartifactsfeature. Saved pages and the Artifacts tab of the dashboard keep working.ori aliaswrites executables by default. A bareori aliasnow writes native-named executables such as~/.local/bin/claude, as--binarydid before. Useori alias --shellfor the shell alias block, which is nowalias claude='ori claude --'. The shell form asks before it edits the startup file, and--yesanswers for scripts.ori auth biometricremoved. Theori auth biometricsubcommands, the desktop credential broker, and the MCP passphrase prompt in/mcpare gone. Saved credentials now live in one encrypted store, as the Sign-in and credentials section describes.- Telemetry content variables removed.
ORI_CLAUDE_OTEL_LOG_CONTENT,ORI_CODEX_OTEL_LOG_USER_PROMPT, andORI_OMP_OTEL_CAPTURE_MESSAGE_CONTENTno longer have an effect. UseORI_TELEMETRY_LEVELinstead. - Feature route access is enforced. A route with no
access, or withaccess: "internal", now answers401without the daemon credential. Aprivateroute needs the credential or a signed browser session, and the Slack dashboard is now aprivateroute. publicroutes are deprecated. Apublicroute still works, but a project feature that declares one gets the boot warningORI_BOOT_REGISTRY_API_ROUTE_PUBLIC_DEPRECATED. Declare a webhook underapi.webhooksinstead.- Eval files must be ES modules.
ori evalno longer loads an eval file that usesmodule.exportsorexports.x. Replacerequire("bun:test")withimport { test, expect } from "bun:test". A file that usesrequireonly for other modules keeps working. - New
costPickshape. Inori eval --json,data.conclusion.costPickis now{ model, passed, costUsd, against, settleAtRuns }. Theratio,cheaperUsd, anddearerUsdfields are gone. - Eval error names. An Ori assertion failure now prints as
EvalAssertionError (ori 1a2b3c4d): ..., with a suffix.instanceofchecks still work, but code that compareserr.name === "EvalAssertionError"no longer matches. - Default reasoning effort. A daemon turn whose
parametersdo not setreasoningnow gets the default effort of the model, oftenhigh. Before, it sent no effort. Sendparameters: { reasoning: {} }to keep the choice of the provider. A turn with noparametersis unchanged. - OTLP span attributes renamed. Spans no longer carry
ori.model,ori.openrouter.session.id,ori.tool.name,ori.model.ttft_ms, orerror.kindonchatspans. Readgen_ai.request.model,gen_ai.conversation.id,gen_ai.tool.name,gen_ai.response.time_to_first_chunk, anderror.typeinstead. Log records keep their names. - Markdown lint.
ori lintandori ci lintnow lint Markdown with rumdl. A project.markdownlint.jsonis no longer merged on top of the framework config.
Codex
- No OpenAI traffic at telemetry off. At
ORI_TELEMETRY_LEVEL=off,ori codexandori harness setup codexturn off Codex analytics, plugins and apps, and setup hides a saved ChatGPT login from plaincodexand Codex Desktop the wayori codexalready does. Codex then sends no metrics toab.chatgpt.com, no featured-plugins, MCP Apps or settings request tochatgpt.com, no token refresh toauth.openai.com, and no plugin marketplace sync to GitHub, inori codex, plaincodex, and Codex Desktop after setup. Codex plugins and apps are off at that level, the saved login stays on disk unused, and acodex loginlasts only for that process.ori harness reset codexremoves all four settings. The forms that run on Codex’s shared background server (agents,queue,remote-control,app-server daemon,--remote) read onlyconfig.toml, so at that level they stop withORI_CLI_CODEX_DAEMON_NEEDS_SETUPuntilORI_TELEMETRY_LEVEL=off ori harness setup codexhas written them. A levelori codexandori harness setup codexcannot read, such asOff, now stops them withORI_CLI_TELEMETRY_LEVEL_INVALID, naming where it was set, instead of quietly counting asno-content. A machine set up on 0.15.7 keeps its old setup afterori update, which does not rewriteconfig.toml; at telemetry offori doctor codexnow reports that setup as a conflict, andori codexand the firstoricommand after the update print one line, each namingORI_TELEMETRY_LEVEL=off ori harness setup codexas the fix. - Nothing written at launch.
ori codexdoes not writeconfig.toml, the thread index or the model catalog in your Codex home. The OpenRouter session ID it keeps for the Codex home now lives under~/.ori/codex. An ID an older release left in the Codex home is read and copied there, so no thread changes provider, and the old file stays. - One provider for setup and launch.
ori harness setup codexnow writes theori_openrouterprovider thatori codexuses. Plaincodex,ori codex, and Codex Desktop list the same threads in/resume. Running setup again moves anori_openrouter_setupsetup to the new form. - Every saved login runs on OpenRouter. With a saved ChatGPT login on any plan, an API-key login, or
CODEX_API_KEY,ori codexruns on OpenRouter and hides the saved login.auth.jsonstays unchanged. The local relay and the Free-plan notice are gone. - Older plain Codex threads.
ori codex resume <id>andori codex resume --lastcontinue a thread that plaincodexmade before Ori, on OpenRouter. - Tool search on by default.
ori codexand the catalog that setup writes turn on Codex tool search for every model. Codex then loads MCP tools andspawn_agenton demand. Administrators can turn it off with theORI_CODEX_TOOL_SEARCHlaunch setting. - Daemon-backed commands. After setup,
ori codex -- agents,queue,remote-control,app-server daemon, and--remoterun through OpenRouter and receive the Codex settings of the workspace. Without setup, they fail withORI_CLI_CODEX_DAEMON_NEEDS_SETUPand say what is wrong with~/.codex/config.toml. - Review and subagent models. A native
review_modeloragents.default_subagent_model, such asgpt-5.4-mini, now routes to its OpenRouter ID, such asopenai/gpt-5.4-mini. This works inori codexand after setup. A caller-c review_model=...still wins. - One session ID per Codex home.
ori codexsends the sameX-Session-Idacrossexec,exec resume, andexec forkon one Codex home, and setup writes that ID into the provider headers. When Ori cannot read the ID file under~/.ori/codex, or it holds no usable ID, the launch fails withORI_CLI_CODEX_SESSION_ID_UNAVAILABLE. When Ori cannot write it, for example on a read-only~/.ori, the launch still starts with the old or a new ID, prints one warning, and threads may move between providers. - Launch matches setup.
ori codexrequests now carryHTTP-Referer: https://or.botandX-OpenRouter-Title: Ori. Websockets are off by default, and-c model_providers.ori_openrouter.supports_websockets=trueturns them on. The/modelpicker lists the model of the run first. - Network loss. When OpenRouter becomes unreachable,
ori codexstops afterReconnecting... 5/5and exits with the connection error. To keep retrying, pass-c features.unbounded_connection_retries=true. - Late web search answers. In a workspace that blocks web search, the first turn now completes even when the policy check answers late. Ori saves the key as blocked, so the next launch leaves
web_searchout. - Faster model list.
ori codexwaits up to 3 seconds for the models of your key, instead of 10. On a timeout, it keeps the catalog of an earlier launch. - No models from another key.
ori codexkeeps one launch catalog for each key. When the catalog fetch fails for a new key, Codex shows its own list instead of the models of the previous key. - Version check in setup.
ori harness setup codexnow refuses a Codex CLI below 0.118.0, asori codexdoes, and leaves~/.codexuntouched. - Large MCP servers on short model names. When Codex asks for a short OpenAI name such as
gpt-6-sol, for example in/reviewor Codex Desktop background jobs, Ori no longer turns on Codex code mode for it. Before, code mode put every MCP tool into one very long description, and a large MCP server could push the request past the size limit and break the session. - Picker names and descriptions. In the Codex model picker, an alias model is named
<alias> · <target name>. Descriptions are cut to 140 characters at a sentence or word end.
Codex setup and Codex Desktop
- Codex Desktop on Windows. On Windows,
ori harness setup codex,ori harness reset codex, andori doctor codex --fixquit a running Codex Desktop, apply the change, and open it again. - Windows lookup failures. When Ori cannot read the Codex Desktop package and
ChatGPT.exeis running, setup stops withORI_CLI_HARNESS_ALREADY_RUNNING. When the app is not running, setup continues and tells you to restart Codex Desktop if it is open. - Windows machine config. On Windows, setup and reset read the system Codex config from
%ProgramData%\OpenAI\Codex\config.toml, as Codex does.ori doctor codexlists that file andrequirements.toml. Ori no longer looks formanaged_config.tomlon Windows. - Desktop model picker. The first 100 catalog entries are now all models that Codex Desktop can show. Setup ranks the setup
--modelfirst, then themodelinconfig.toml, the models of Codex, and the newest OpenRouter models. Older models stay usable throughconfig.toml,--model, or/modelin the Codex CLI. - Background tasks in guardrailed workspaces. The catalog no longer holds hidden entries for models that the key cannot use. In a guardrailed workspace, Codex Desktop titles, summaries, and commit messages run on the catalog default instead of failing.
- Catalog refresh. While Codex is open, the setup catalog follows the models of your key within about 5 minutes. Codex shows the new list on its next start. The detached refresh now stops after 10 seconds, instead of 30.
- Setup receipt. Setup records what it wrote in
$CODEX_HOME/ori-openrouter-receipt.jsonand marks itsconfig.tomllines with# ori harness setup. A failed setup rerun leaves the earlier setup as it was, including the thread index and the transcripts. - Inline tables. Setup now works with an inline
features = { ... }oragents = { ... }table that has other members. Before, it failed withORI_CLI_HARNESS_CONFIG_INVALID. - Locked thread index. Setup and reset open the thread index of Codex again after an
SQLITE_CANTOPENerror, and they wait out a lock that Codex holds. - Unreadable thread index. When the index stays closed, a setup rerun and
ori doctor codex --fixstill finish. They keep the earlier provider of each thread in~/.codex/ori-openrouter-state.json, so a laterori doctor codex --fixcan move the threads back. - Reset keeps its record. When reset cannot rewrite a transcript, it keeps
~/.codex/ori-openrouter-state.jsonand tells you to runori doctor codex --fix. Athreadstable withoutrollout_pathstops reset with an error that names the table and the column. - Blank lines kept. Setup and reset keep the trailing blank lines of
config.toml. - Doctor after a moved binary. When the
oribinary moved,ori doctor codex --fixpoints the credential helper at the new path. It also reports when an older Ori wrote the setup. - OpenAI-hosted plugins.
ori harness setup codexnow says that OpenAI-hosted plugins need a ChatGPT login and do not work through OpenRouter. Local plugins and skills keep working. AtORI_TELEMETRY_LEVEL=off, where setup turns Codex plugins off, it says instead that plugins, local ones included, are off and that skills keep working.
Desktop apps
- Desktop app install. On macOS,
ori harness setup claudeandori harness setup codexoffer to install the latest desktop app when it is missing or too old. The offer appears only when Ori can prompt. - Install checks. Ori downloads the app only over HTTPS from the vendor host, checks the bundle ID, and runs a Gatekeeper check before it copies the app. Ori installs into
/Applications, or into~/Applicationswhen/Applicationsis not writable. A failed swap puts the old copy back. - Running-check errors. When Ori cannot check whether Codex Desktop or Claude Desktop is running, the error names the failed command, for example
/usr/bin/pgrep exited 2. On Windows, the Codex Desktop check stays under one second whenPSModulePathis missing.
Claude Code
- Background session commands.
ori claude attach,logs,stop,kill,rm, andrespawn <id>now run the matching Claude Code commands. They accept a short ID, a full session ID, or a name. - Other subcommands.
ori claude mcp,auth,plugin,doctor,update, and the other subcommands reachclaudewithout the arguments that Ori injects.ori claude agentsstill routes dispatched sessions through OpenRouter. - Claude Code accounts.
ori claude auth ...andori claude setup-tokenmanage the real account store of Claude Code, so plainclaudesees the same login. - Background sessions keep their login. A background session answers on the login that it launched with, also after
claude attachor a daemon restart. When an environment-only key is gone, the credential helper refuses instead of using another account. - Tool search for Anthropic models.
ori claudenow setsENABLE_TOOL_SEARCH=truewhen every launch model is an Anthropic model. It setsENABLE_TOOL_SEARCH=falsefor any other model, such as--model openai/gpt-5.5. Your ownENABLE_TOOL_SEARCHstill wins. Administrators can turn the default off with theORI_CLAUDE_TOOL_SEARCHlaunch setting. - Version gate names your setting. When a workspace raises
ORI_CLAUDE_MIN_VERSIONor lowersORI_CLAUDE_SUPPORTED_BELOW, the refusal names that setting. A prerelease such as2.1.116-beta.1counts as older than 2.1.116. - Claude Desktop 2.31226.
ori harness setup claudeconfigures Claude Desktop 2.31226.0 and 2.31226.1 again. Before, setup skipped them withfails Ori's setup contract checks: credential-helper-protocol. - Claude Desktop 3.x and later.
ori harness setup claudeconfigures Claude Desktop 3.x. Ori no longer refuses a Claude Desktop version because of its major version. A version newer than the verified one gets a warning, and setup configures it only after it passes the setup contract checks. A failed check still refuses. - The copy macOS opens. On macOS, setup first asks LaunchServices which copy of Claude Desktop macOS opens, and checks that copy. Before, setup could pass on any copy that Spotlight found, such as an old backup that you do not run. When the copy macOS opens is too old and another copy is in range, setup skips Claude Desktop and names both paths.
- Exact reset.
ori harness reset clauderestoresclaude_desktop_config.jsonbyte for byte when only thedeploymentModevalue of Ori changed. Reset also gives back anavailableModelslist that setup replaced, unless you changed it since. - Only the entries that setup added. Reset removes only the
WebSearchentry that setup added topermissions.deny, and your own entries stay. A catalog refresh that fails partway leavessettings.jsonand the setup records as they were.
Pi and other launchers
ori openwork. This command launches theopenworkexecutable of the OpenWork desktop app through OpenRouter.--modelsets the default model.--reasoning-effortis refused, because OpenWork picks variants in its own UI.ori antigravity. This command launches the Antigravity CLIagythrough OpenRouter with your Ori credential.--modelpasses the model toagy. Whenagyis missing,ori harness listshows its install command.- Antigravity settings. Administrators can set the Antigravity sandbox, terminal, file access, and review settings. Ori merges them over
~/.gemini/antigravity-cli/settings.json. - OpenWork details. OIDC sign-ins get a credential plugin, so long sessions survive token expiry. Ori warns when OpenWork’s AI Providers settings hold another OpenRouter key, and leaves that key alone. On macOS and Windows, link the app executable onto
PATHasopenwork. - Rejected keys stop the launch. When OpenRouter rejects the key with a 401 or refuses it with a 403,
ori claudeandori codexstop withORI_CLI_OPENROUTER_KEY_REJECTEDbefore the harness starts. A 401 or 403 without the error body of OpenRouter prints a proxy warning, and the launch continues. - OIDC sessions refresh once. When OpenRouter refuses an OIDC session, the launch refreshes the session once and tries again. If that fails, the launch tells you to run
ori login --oidc. An OIDC launch no longer falls back to the public model list. - Shared version refusals. Claude Code, Codex, and Pi version refusals now share one message. It names the found and the needed version, and the hint gives the upgrade command.
- Pi attribution.
ori pirequests now carryHTTP-Referer: https://or.botandX-OpenRouter-Title: Orieven when Pi sets its own values. They no longer carryX-OpenRouter-Categories: cli-agent. - Pi startup models.
ori pi, and plainpiafterori harness setup pi, start with the models of your key even when the credential helper is slow. A shared 3 second budget now capsORI_PI_STARTUP_CATALOG_BUDGET_MS. - Pi prereleases.
ori pinow refuses a prerelease of the minimum version, such as0.84.0-rc.1. A raisedORI_PI_MIN_VERSIONis named as the reason. - Pi reset on Windows. On Windows,
ori harness reset pinow removes~/.ori/pi-setupwhen no other agent folder uses it. ori prime-agentstartup. The startup model request now stops after 3 seconds. On a timeout or a failure,ori prime-agentkeeps its built-in model list and prints the existing warning.- Windows project folders. On Windows, a project folder is free again as soon as Codex or Claude Code exits.
- Temporary folders. A harness settings preflight that times out no longer leaves an
ori-harness-isolated-*folder in the temp directory.
Administrator controls
- Native harness settings. Administrators can now set native Claude Code, Codex, and Pi settings as
claude.<key>,codex.<key>, andpi.<key>. The values come from launch settings, managed preferences, or/etc/ori/config.json. Nothing changes until an administrator sets one, and a JSON value is written as JSON text. - Claude Code settings. Ori offers 141 Claude Code settings. It passes 133 through
--settingsand enforces 8 through--managed-settings. The catalog lists 55 more as unavailable, with the reason. - Codex settings. Ori offers 67 Codex settings, and the catalog lists 37 more and every
requirements.tomlkey as unavailable. When Codex refuses a value, Ori launches without it and prints one line that names the setting. - Pi settings. Under workspace Pi settings,
ori pistarts on an overlay agent folder under~/.ori/pi/agent-overlays/. Itssettings.jsonholds your values with the workspace values on top. Ori saves your own changes back into yoursettings.json. - Pi precedence. A Pi flag that you pass, such as
--thinking, beats the workspace. A project.pi/settings.jsonstill overrides the workspace inside that project. When Pi would refuse a value, Ori drops only that value and prints one line. - Pi value rules. Ori checks Pi values against the Pi 1.2.0 settings schema. For example,
pi.httpIdleTimeoutMsaccepts"disabled"and refuses-1. - One model family. With
ORI_CLAUDE_CROSS_FAMILY_MODELSoff,ori clauderuns onlyanthropic/*models and refuses any other model before Claude Code starts. WithORI_CODEX_CROSS_FAMILY_MODELSoff,ori codexruns onlyopenai/*models. The/modellists and the desktop setup catalogs follow the same rule. - Codex image generation. With
ORI_CODEX_IMAGE_GENERATIONoff,ori codexdoes not offer image generation, and a caller flag cannot turn it back on. Codex Desktop follows after the nextori harness setup codex. - Telemetry level.
ORI_TELEMETRY_LEVELtakesfull,no-content, oroff, and defaults tono-content. Atoff, Ori exports nothing, every harness starts withOTEL_SDK_DISABLED=true, and anonymous product telemetry is off. - Workspace collector.
ORI_TELEMETRY_COLLECTORnames the collector for Ori and harness telemetry. Only an administrator can set it, so a cloned repository cannot send your prompts to its own collector.ori doctorreports the level, the destination, and whether it carries content. - Content stays off the laptop collector. Content never goes to the local
OTEL_EXPORTER_OTLP_ENDPOINTcollector, even atfull. Ori removes content variables such asOTEL_LOG_USER_PROMPTSfrom every launch that carries no content. - Organization minimum version.
ORI_MINIMUM_VERSIONsets the oldest Ori that the organization allows. Below it, every command exceptdoctor,help,update, andversionfails withORI_CLI_VERSION_BELOW_ORGANIZATION_MINIMUM. A value in the shell or a user file is ignored. - Allowed commands.
ORI_ENABLED_COMMANDSlists the top-level commands that may run, for examplelogin,codex,pi. Other commands fail withORI_CLI_MANAGED_COMMAND_DISABLED.helpandversionalways run, and a launcher must also passORI_ENABLED_HARNESSES. A value in the shell, a user file, or a workspace file is ignored. - Update below the minimum. Below
ORI_MINIMUM_VERSION,ori updateruns even whenORI_ENABLED_COMMANDSleaves it out. - Published minimum notice. On managed installs, the published-minimum notice now prints before the command, so
ori codexandori claudeshow it. OnlyORI_NO_UPDATE_CHECKhides it. UnderORI_DISABLE_UPDATES, the notice tells you to ask IT. - Doctor version row. The Configuration section of
ori doctorstarts with a Version row. It names the running release, the organization minimum, and the published minimum.
Sign-in and credentials
- Gateway question. In an interactive browser sign-in,
ori loginand the sign-in of a harness now askWhich OpenRouter gateway should Ori sign in through?, with Global first. When you choose US, Ori exchanges and checks the key onus.openrouter.aiand savesregion: "us". - When Ori skips the question. Ori does not ask when
--region, a saved region,ORI_OPENROUTER_REGION, a trustedORI_OPENROUTER_BASE_URLorORI_OPENROUTER_ENDPOINT, or a managedORI_REQUIRED_REGIONdecides the gateway. Theori auth regionpicker now lists the configured region first. - Remaining budget.
ori authprints abudget:line with the amount left, the limit, the scope, and the reset time.ori auth --jsonadds abudgetobject.ori doctorshows the same line and warns at 80% used or when the budget is spent. - Budget notice at launch.
ori claude,ori codex, and the other launchers print one line when 80% or more of the budget is used. Ori never refuses a launch for this. WhenORI_HIDDEN_FIELDSnamescreditLimit, Ori shows no budget anywhere. - Spent budget in
ori code. A budget refusal inori codefails withORI_OPENROUTER_BUDGET_EXHAUSTED. The message names the scope, the figures, and the reset time, and suggests a free model. - One credential store. Saved API-key logins, OIDC sessions, and MCP grants now live in
~/.ori/auth/credentials.sqlite. UnderORI_CREDENTIAL_STORAGE=protected, the key is the keychain itemdev.openrouter.ori.credentials.v1. Underfile, the key is~/.ori/auth/credentials.key. - Automatic import. The first command after the upgrade imports the earlier credential layout and removes it. When an administrator chose
fileand the old layout used the keychain, the import refuses withORI_CLI_CREDENTIAL_IMPORT_DOWNGRADE_REFUSEDand namesori auth migrate --yes. ori auth migrate. This command imports an earlier layout, moves the key into the custody that the policy names, and promotes a plaintext JSON login. A move intofilecustody needs--yes.- Custody mismatch. When the policy names one custody and the key is in the other, commands that use a saved credential fail with
ORI_CLI_CREDENTIAL_CUSTODY_MISMATCH.OPENROUTER_API_KEYkeeps working. - Keys that were readable on disk. After an import from a plaintext layout,
ori authtells you to runori loginto replace the key.ori auth --jsonreportsexposedOnDisk: trueuntil you sign in again. - Sign out everywhere.
ori logout --allsigns out every saved login and OIDC session on the machine and keeps the store.ori auth resetnow only discards a store that cannot be read, with its key. - Storage in status.
ori authandori doctorprint onecredential storage:field that names the provider and who chose it.
CLI output
- One output style. Every command now prints results, notes, warnings, and failures with the same marks:
✔,◇,⚠, and✗, with a↳hint below. Lines wrap at the terminal width under their mark. Piped output does not wrap. - Color. The marks are colored on a terminal.
NO_COLOR=1,FORCE_COLOR=0, andTERM=dumbturn color off, andFORCE_COLOR=1turns it on. - Help.
ori --helpshows the commands in five groups. Every command page uses one layout withUsage,Arguments,Flags,Details, andExamples. A bare group, such asori workspace, prints its help and exits 2. - Help text. Launchers read
Launch <product> through OpenRouter. Ori corrects help that was wrong, and a launcher that cannot translate--reasoning-effortsays so on its own page. - Machine output. Under
--json, stdout carries only the envelope, and a failure also prints its✗line and hint on stderr.ori login,ori harness setup,ori harness reset,ori pack,ori link,ori unlink,ori profile, andori aliasnow print envelopes in machine mode. - When Ori asks. A prompt opens only in human mode, with stdin and stdout on a terminal and
CInot set to a true value. Otherwise the command takes its non-interactive path, or fails withORI_CLI_TERMINAL_REQUIREDand names the reason. For example,CI=1 ori initno longer opens the name prompt. - Confirmations. A confirmation defaults to the safe answer and names the flag that skips it, such as
--yes. After Ctrl-C at a confirmation,Cancelled.starts on its own line. - Sign-in URL on stderr.
ori loginprints its sign-in URL and code on stderr, soori --json loginkeeps stdout for the envelope. A non-interactive login no longer offers the prompt to paste a code. - Reports and totals.
ori doctor,ori features validate,ori features gate,ori mcp list,ori mcp test,ori ci, andori update --checkend with one totals line.ori harness listmarks a missing launcher with◇and its install command, not✗. - Steps.
ori init,ori pack,ori login, dependency installs, and other waits show each step with its time.ori initends with one summary and the next commands to run. - Tables.
ori sessions,ori logs ls,ori daemon list,ori schedules list,ori interns list,ori skills list,ori features list,ori operations, andori profileprint aligned tables. A narrow terminal drops the least important columns first, and IDs are never cut. - Runtime lines.
ori start,ori dev, andori logsprint each runtime line as a time, a scope, and a message, such as14:02:11 ◇ runtime listening on http://127.0.0.1:53211. These lines always go to stderr.ori logsshows the stored time of each record. - Empty states. A command with nothing to show prints a
◇line and a hint. For example,ori logsprints◇ No runtime log entries yet, andori sessionsprints◇ No sessions recorded yet. - Quiet launches. A bare
ori claudeno longer prints linked sessions before Claude Code starts. Credential conflicts and version refusals print in one shape with product names, such asOpenCode,Oh My Pi, andFactory Droid. - Product names in JSON.
ori harness list --jsonnow reports thedisplayNamevaluesOpenCodeandOh My Pi, instead ofopencodeandomp. - Codex app. Messages, hints, and help pages now call Codex Desktop “the Codex app”.
- Install offer. When a product is not installed, the launcher asks to run the install command and defaults to no. Without a terminal, it prints one
✗line with the quickest install command. - Setup and reset narration.
ori harness setupprints what it routed and what it replaced, and ends with↳ Run `ori harness reset <kind>` to undo this.ori harness resetprints what it put back, or says that there is nothing to reset. - Safe text. Ori replaces control characters, line separators, and bidirectional overrides in printed paths, model IDs, and provider messages with spaces.
- Wrapped failures. When one command wraps the failure of another, the envelope
codeis the inner code, and the hint survives. - No raw boot logs.
ori features validate,ori harness list,ori init, andori updateno longer print rawERRORboot records above their reports.ori harness listwarns when features failed to load, and its JSON addsfailedFeatures. - Environment flags. Ori reads
CIand similar flags the same way everywhere, so" 0 "and" false "count as unset. - New failure code.
ori initwith no name and no terminal now fails withORI_CLI_PROJECT_NAME_MISSING, instead ofORI_CLI_WORKSPACE_CREATE_FAILED.
CLI
ori unalias.ori unaliasremoves the binary aliases, then the shell alias block.--binary,--shell, or--shimremoves one kind only.- Alias checks in doctor.
ori doctoradds an Aliases section. It flags a binary alias with no real binary behind it, an alias that a real binary shadows, and a shell alias without--. - Workspace reset and the code daemon.
ori workspace resetnow refuses withORI_CLI_WORKSPACE_RESET_DAEMON_RUNNINGwhile anori codedaemon runs. Stop the daemon withori daemon stop <id>, then run the reset again. - Update prompts. Only
ori codechecks for updates and asks to install them. Not now is remembered for that version. Other commands make no update check, but can still print a cached notice. ori updateand the code daemon. After installing a release,ori updatestops an idleori codedaemon, so the nextori codestarts the new build. A busy daemon keeps running, andori updatetells you to runori code daemon stopwhen it finishes.- A daemon that is shutting down.
ori codenow starts a fresh daemon when the running one is shutting down, and no longer attaches to the old one. ori devpackage changes.ori devnow reports a feature package change that you make right after startup.
Chat TUI
/modelwith a near name./model opus55,/model opus 5.5, and/model gpt-5-miniswitch to the model that they name and printModel set to <id>. A name that matches several models opens the picker, filtered to that name. An unknown name opens the picker and keeps the current model.- Fuzzy picker search. In the
/modelpicker, typinggpt41oropus 5.5findsopenai/gpt-4.1oranthropic/claude-opus-5.5. - Picker sort. In
/model, Ctrl+O cycles the sort through popular, newest, A-Z, and provider. The header shows the mode when it is not popular. - Decisions and retries. The collapsed agent action now holds model decisions and retry progress, with the attempt number, the wait before the next try, and the result. Turn and runtime failures appear in the same group, not in a separate card.
- Local MCP servers.
ori codecan call tools of astdioMCP server without a network grant. In/mcp,stdioservers readLocal, and HTTP servers without OAuth readNo sign-in needed. - Decision details. Press Ctrl+O on a recovered agent action to see the options of each decision with score bars. A retry row says
by Jevorby policywhen a decision record shows what chose the retry. - Separator notices. Steering, steering fallbacks, and discarded queued messages now show as separator lines in the transcript.
- Light theme. Pickers, permission prompts, and scrollbars stay readable on the OR Light theme.
- Temporary files. The compiled
ori codeandori tuiremove their tree-sitter temp files when they exit.
Slack
- Streamed replies. Under
agent_view, the reply opens on the first output, and the answer streams as the model writes it. The agent session of the thread shows processing while the intern works. The status line and the rotating loading messages are gone. - Tool steps in a plan. Each tool call is a task card in one plan block, with the tool name and a short public summary. Ori redacts secrets in the summary. The plan title counts the steps, for example
Working · 3 of 5 steps done. - Acknowledgement first. The intern writes a one-sentence acknowledgement before its first tool call. When the turn ends, the text after the last tool call becomes the answer below the plan.
- Stream failures. When Slack refuses a stream call, the intern posts the full answer as a normal message when the turn ends. In rare cases, the answer appears twice. A retried attempt closes its stream and posts its answer at the end of the turn.
- Retries and decisions. A scheduled retry shows as a plan step with the attempt number and the wait before the next try. Each model decision shows as a finished step, and See details lists it.
- Slack’s Stop. Selecting Stop in Slack ends the running turn as cancelled, as
!stopdoes. The intern no longer posts a Working message with a Cancel button. The Slack app must subscribe toagent_session_stoppedfor Stop to work. - Stop setting. The dashboard setting “Let Slack’s Stop button stop the intern” replaces “Show Cancel button in Slack” and is on by default. With the setting off, Ori ignores Stop.
- Who can use the intern. A new dashboard section, “Who can use this intern”, lists the people that the intern answers. With people listed, the intern ignores messages, buttons, forms, and Stop from anyone else.
SLACK_ALLOWED_USER_IDSstill works and adds to the list. - Auto-responders in every channel. An auto-responder card can now answer in “All channels the intern is in”. This covers public and private channels, but never direct messages. A card for specific channels wins in those channels.
- Reply instructions. The dashboard has an editable Reply instructions section of up to 8,000 characters, with Reset to default. The default reply prompt is shorter, and every turn gets the same text.
- Dashboard design. The dashboard now matches the OpenRouter intern page, with Behavior, Auto-responders, and Threads sections. Settings are switches, and threads are a table with status badges and controls. A refused save returns an HTML page with a link back.
- Thread commands. A mentioned command, such as
@intern !status, now posts one reply instead of two.!qand!steeralways reach the intern, also in a muted thread. A command in another channel is no longer taken for a duplicate. - Spawned threads. The origin thread shows “Spawned a new thread” with a View spawned thread button, instead of an escaped link.
- Settled turns. The session of the thread returns to active after every finished, failed, cancelled, or interrupted turn. Before, it could spin until the one-hour timeout of Slack.
- Old threads. A reply in a thread whose working notes are gone now gets the notice and a fresh session at once, without a failed turn first.
Interns
- Chores. Ask the intern to “take this off my plate”, and the schedule skill loads. When you keep asking for the same safe work, the intern offers to take it over on a schedule. It does not offer this for merging, deploying, or messaging people.
- Memory example.
examples/features/memoryis an opt-in feature that gives an internmemory_save,memory_recall, andmemory_forget. It keeps memories as plain Markdown undermemory/log/. Its guidance says when to use memory and when to change a skill, schedule, or feature instead. - Feature change history. The History of the intern page lists feature adds, edits, and removals. Each entry links to the chat session that made the change, when Ori can trace one. The report carries feature names, never file contents.
- Previews.
preview-statusand preview warm-up now reach the repositories that a GitHub organization grants.preview-statusadds aConnected, not warmed yet:line. ori atand/resume. Afterori at <intern>,/resumelists the persisted sessions of the intern instead of loggingInvalid URL.- Cost pages. Ori stores each model call when the call is reported, so intern cost pages show less “Other usage”.
- Saved pages. Saved web pages carry
worker-src 'none', so a page cannot start a worker from adata:URL. - Feature
#imports. Interns and other compiledoribinaries can load features whosepackage.jsondeclaresimports. A#import with no matching entry fails the feature build with a message that names the file.
Workflows and feature authoring
- Durable workflows. A feature can ship workflows in
workflow.ts, inworkflows/<name>/workflow.ts, or as aworkflowexport offeature.ts. Build them withdefineWorkflowfromori, usingctx.step,ctx.fanout,ctx.call, andctx.sleepwith an explicitkey. ori workflows.ori workflowslists the loaded workflows, andrun,runs,show,cancel, andresumemanage runs.show --mermaiddraws a run.run --waitexits non-zero withORI_CLI_WORKFLOW_RUN_FAILEDorORI_CLI_WORKFLOW_RUN_CANCELLED.- Workflow routes. The daemon serves workflows under
/api/workflows, and every route needs the daemon credential. WithORI_WORKFLOW_RESUME_ON_BOOT=off, the daemon does not resume interrupted runs at boot.ori skills get workflowexplains authoring. - Required connections. A feature can declare
ori.requires.connectionsinpackage.json, such as["github", "linear"].ori features addshows each connection as granted, missing, or not checked, and refuses when one is missing.GET /api/featuresincludesrequires.connections. - Private routes in the daemon API.
GET /api/features/<id>/routes/<path>serves theprivateGET routes of a feature with the daemon credential. Skill scripts readORI_DAEMON_TOKEN, or~/.ori/daemon/<port>.tokenwhen the variable is unset. - Runtime address. After
ori start --port <port>binds, the daemon environment holdsORI_RUNTIME_HOSTandORI_RUNTIME_PORT, so schedule handlers and helpers can read them. - In-run spawns at the limit. A tool that spawns an unparented thread while all 64 run slots are busy now gets a failed child with
ORI_HISTORY_CAPACITY. A spawn on the session of the tool fails withORI_SESSION_RUN_IN_PROGRESS. Before, both runs never ended. - Tool calls without a name. When the model sends a tool call with no name, the native harness retries once and then answers. Before, the turn failed with
ORI_MODEL_REQUEST_INVALID.
Daemon API
- Session conversation.
GET /api/sessions/:id/conversationnow includes reasoning, tool starts and failures, toolinput,result, anddurationMs, decisions, and retries.turn.startedcarries the names of the attachments. The route also answers faster. - Session history check.
GET /api/sessions/:id/historyanswers200while the history store holds the session, and404once it does not. - Paged summaries.
GET /api/sessions/persisted?view=summaryacceptslimitandoffset, andGET /api/sessions/<id>?view=summaryreturns the summary of one session. A summary keeps a bounded set of runs and setsrunsOmitted: truewhen it leaves origins out. - Feature install.
PUT /api/features/:featureId/installinstalls a feature folder, andDELETEon the same path removes it.
Evals
ori eval prompt. This new command compares models on prompts without an eval file. Pass prompts with--promptor a YAML suite, two or more--modelslugs, and an optional--judge. It writesprompt-eval-report.mdin the current folder.- Checks before spending.
ori eval promptprints a plan line before it runs.--dry-runchecks every model and the judge against your key and calls no model.--pilot 1runs one prompt on every model to price the full run. - Hidden criteria. Every prompt eval runs hermetic, and the generated eval deletes its criteria before any candidate runs. The default judge is
~anthropic/claude-opus-latest. - Responses in the report. Every
--reportnow ends with a## Responsessection that shows the answer, the tools, and the reason of each judge for every run.ori eval --jsonaddspassReasonsto passing judged runs. - Fair comparisons. When a model errored on a case, or the judge could not grade it, Ori leaves that case out for every model. The conclusion lists the left-out cases, and
--jsonlists them indata.conclusion.droppedSlots. - Numbered attempts.
agent.runacceptsattemptto number the repeats of a case, so Ori leaves out only the repeat that failed. - Cheapest close model. The conclusion names the cheapest model that the runs could not tell apart from the best, with pass counts and costs. It also says how many runs would tell them apart. Only models with a reported cost above $0 count for the cost pick.
- No hidden fallbacks. In evals, a failing candidate or judge now shows as failed, whatever
ORI_FALLBACK_MODELSsays. PassfallbackModelsinparametersto test a fallback on purpose. - Models your key can use.
candidateModels,rankedModels,isModelLive,assertModelIsLive, andcandidateEffortsinori/evalnow return only models that the key can call. Under a key, they fail closed when Ori cannot confirm the models of the key. Ori checks the default judge against the key too. - Failed loads. An eval file that does not finish loading now declines the conclusion and fails the run.
data.scope.filesUnloadednames the file. - Baselines after a failed load.
--baseline lastand--baseline bestno longer report a falsepass-rateregression after a run where an eval file failed to load. - Failed tests. A failed test whose error no run recorded now blocks the winner. For batched cases, end the test with the new
throwCaseFailures(failures)export. - Clearer messages. When
ori evalstops calling a model, it names the last failure code, and--jsonrows addoutcomeCode. Multi-line run messages are indented under their run. The--pilotline says whether Ori recorded a full-sample pilot. - Report on screen.
ori evalprints a report that names the pick, the totals, and each model that is out. Warnings go to stderr as⚠lines.
Telemetry
- Schedule traces. The run of a schedule fire now joins the trace of the fire.
- Failed model calls count. A failed model call records the tokens that its response reported.
- Retried sends. A failed telemetry send is retried on the next flush instead of lost.
Harness versions
- Verified versions. This release verifies Claude Code 2.1.299, Claude Desktop 2.37337.1, Codex CLI 0.162.1, and Codex Desktop 26.1007.71326.
ori codex and Codex Desktop can now generate images through OpenRouter, and context compaction works in ori codex relay mode. ori harness setup now covers Claude Code, Claude Desktop, Factory Droid, and Pi, and ori profile keeps the harness sessions of each profile apart. Administrators can pick the allowed harnesses, cap the key limit, and serve workspace settings from OpenRouter. Interns gain web page artifacts and Slack thread commands, and shell and eval runs fail less often.Existing OIDC sign-ins keep working for inference in every harness. Only ori at and ori interns attach ask you to sign in again with ori login --oidc, to grant the new intern access.Codex
- Image generation. Image generation is on by default in
ori codexand in Codex Desktop afterori harness setup codex. Ask Codex to draw something, and it callsgpt-image-2through OpenRouter. Setup adds"x-openai-actor-authorization" = "ori"to the OpenRouter HTTP headers, andori harness reset codexremoves it. - Free-plan logins. When your saved Codex login is on the ChatGPT Free plan,
ori codexruns the session on OpenRouter without that login, so image generation works. Ori prints a notice that this session does not share/resumewith plaincodex. Other plans keep using the relay as before. - Relay compaction. With a saved Codex login, auto compaction and
/compactnow work inori codex, and the next turn continues from the summary. Before, compaction failed withInvalid Responses API request. - Resume after plain Codex compaction. A thread that plain
codexcompacted against OpenAI now resumes inori codexinstead of failing withinvalid_encrypted_content. - Clear compaction errors. When the summary turn fails, Codex shows the OpenRouter status and message, or the reason the turn stopped.
- Workspaces that block tool search.
ori codexno longer sends theopenrouter:tool_searchserver tool, for any model. A workspace that blocks that tool no longer refuses Codex requests with a 403. - Workspaces that block web search. When the workspace of the key blocks Web Search,
ori codexturns off hosted web search with-c web_search='disabled'. Your ownweb_searchsetting still wins. Ori checks each new key one time with a small request and keeps the answer for one day in~/.ori/codex/web-search-policy.json. - Relay retries without web search. In relay mode, a later web search refusal makes the relay send the request again without web search, so the turn completes.
ori codex resume --laston Windows.--lastfinds the threads of the current folder again. It also works when you enter the folder through a short name or a symlink.
Codex setup and reset
- Desktop plugin search. Codex Desktop plugin search keeps working after
ori harness setup codex. Setup writesfeatures.remote_plugin = false. If you setremote_pluginyourself, setup keeps your value as a comment, and reset restores it. If your setup was made before this release, runori harness setup codexagain to get the plugin search fix. - Doctor flags an older setup.
ori doctor codexwarns when an older Ori wrote the setup, and tells you to runori harness setup codexagain.ori doctor codex --fixleaves that setup as it is and says so. - Your own OpenRouter provider.
ori harness setup codexno longer refuses when[model_providers.openrouter]exists. Setup now uses its ownori_openrouter_setupprovider and leaves your table alone. After reset, your provider is selected again. - Reset finishes the job. When you edit the
model_providerline that setup wrote, reset still removes the line and names it. Before, reset could exit 0 and leave Codex broken. - Doctor finds an undefined provider.
ori doctor codexreports amodel_providerthat no Codex config defines. For an ID that Ori setup used,ori doctor codex --fixremoves the line. - No Codex CLI needed.
ori harness setup codexno longer needs the Codex CLI. It reads the model list from the CLI inside the Codex Desktop app. spawn_agentafter setup. The Codexspawn_agenttool works again for OpenRouter models afterori harness setup codex.- Migration. Existing setups under
openrouterkeep working with reset and doctor. Running setup again moves them, including thread rows and transcript metadata, toori_openrouter_setup. - Codex Desktop restarts for you. On macOS,
ori harness setup codex,ori harness reset codex, andori doctor codex --fixquit a running Codex Desktop, apply the change, and open the app again. A closed app stays closed. The app also opens again when the change fails. - Web search in blocked workspaces.
ori harness setup codexwritesweb_search = "disabled"when the workspace of the key blocks Web Search. The check waits up to 30 seconds. When setup cannot check, it says so and tells you to run setup again if Codex answers 403 for web_search. - Threads move between providers. Setup, reset, and
ori doctor codex --fixrewrite the thread rollouts that the other provider cannot read. Each command prints the number of changed threads and the backup folder under~/.codex/ori-setup-backups/or~/.codex/ori-reset-backups/. To undo a rewrite, copy the backup file back over its rollout. After reset, you can delete the backup folders to free the space. - Final newline kept. Setup and reset keep a
config.tomlwithout a final newline as it was. - Clearer catalog hint. When setup cannot build the catalog, the hint tells you to put
codexon your PATH or install Codex Desktop, and to check that openrouter.ai is reachable.
Claude harness setup
ori harness setup claude. This command configures Claude Code on macOS and Linux, and Claude Desktop on macOS, to use OpenRouter. It has no flags and no prompts. When Claude Desktop is missing, updating, or outside the supported range, setup prints one skip line and configures Claude Code alone.- Managed settings refuse setup. Setup refuses to run when the machine has managed Claude settings, such as
/etc/claude-code/managed-settings.json. - Provider keys move aside. Setup removes
ANTHROPIC_AUTH_TOKEN,ANTHROPIC_API_KEY, andCLAUDE_CODE_USE_BEDROCK,VERTEX, orFOUNDRYfrom the settingsenv, because Claude Code ranks them aboveapiKeyHelper. Setup names each removed key, but never its value. ori harness reset claude. Reset restores the priorapiKeyHelper,ANTHROPIC_BASE_URL,ANTHROPIC_CUSTOM_HEADERS, and Claude Desktop gateway. Reset restores only the keys that Ori still owns. It keeps each key that you changed and names it.- A saved login is required. Setup needs a global or workspace saved login from
ori login. An environment-only or project key fails withORI_CLI_OIDC_LOGIN_REQUIRED. ori claude-desktopis removed. Useori harness setup claudeandori harness reset claudeinstead. No alias remains.- Session IDs after setup. After setup, each
ori clauderun sends its own session ID instead of the install ID. - Settings keep their shape. Setup keeps the indent, line endings, key order, and final newline of
settings.json. When only the keys of Ori changed,ori harness reset claudegives back the original bytes. - Model catalog for plain
claude. Setup writes the OpenRouter models of your key into.claude.json, with OpenRouter prices. Whensettings.jsonhas nomodel, setup selects the Sonnet model of the catalog. /modellists the models of your key. Setup writesavailableModels, so/modelin plainclaudelists only the models of your key. The daily catalog refresh keeps the list current, and it follows a new key afterori login.- Fast mode. Setup sets
CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK, so plainclaudeoffers/fast. Reset removes the value or restores your earlier value. - One-hour prompt cache. Setup writes
promptCacheTtl: "1h"whensettings.jsonhas no value. Reset restores the earlier value while the1hvalue of Ori is still there. - EU region denies WebSearch. When the region is
eu, setup adds a deny entry for WebSearch and says so. Reset removes the entry while it is still in the list. - Empty remote settings. Setup no longer refuses because of an empty
remote-settings.json. The refusal for managed configuration now names the file. WithCLAUDE_CONFIG_DIRset, setup checks the file in that folder.
Claude Code
- One-hour prompt cache.
ori claudesetspromptCacheTtl: "1h"for the main thread. A Claude settings file, a caller--settingsvalue, orCLAUDE_CODE_PROMPT_CACHE_TTLstill wins. Subagents keep the default cache time. To keep 5 minutes, set"promptCacheTtl": "5m"in~/.claude/settings.json. /modelfollows your key.ori claudelimits/modelto the models in the catalog of your key. When the key has no Opus model and nothing chose a model,ori claudestarts on the newest Sonnet, Haiku, or Fable model.- Provider switches are removed. When the shell sets
CLAUDE_CODE_USE_BEDROCK,CLAUDE_CODE_USE_VERTEX,CLAUDE_CODE_USE_FOUNDRY, or another provider switch,ori clauderemoves it and prints one line. Other flags, such asCLAUDE_CODE_USE_POWERSHELL_TOOL, pass through. - Correct launch messages.
ori claudeno longer warns about a:freemodel that the catalog lists. It also accepts Claude Code aliases, such ashaiku,opusplan, andsonnet[1m], without a warning. - Rejected keys and unreachable hosts. A 401 from the catalog prints
OpenRouter rejected this API key. Run `ori login` to sign in again.An unreachable OpenRouter endpoint exits at once withORI_CLI_OPENROUTER_UNREACHABLE.
Factory Droid
ori droid. This command launches the installeddroidwith the OpenRouter catalog, plusopenrouter/auto, ascustom:openrouter/<model id>models. Droid arguments pass through unchanged, but--settingsis refused. API keys and OIDC sign-ins both work, and no raw token is written to a file.ori harness listshows Factory Droid with its install command.- Model routing. On OpenRouter hosts, Anthropic models use the Messages API, so Droid prompt caching applies. OpenAI models use the Responses API, and every other model uses chat completions.
ori harness setup droid. Setup writes the OpenRouter models into the Factory settings, so Factory Desktop can use them after a restart. Your own settings and models are kept, and a model of yours with the same ID wins. Run setup again to refresh the list. Setup needs a savedori loginand runs on macOS and Linux.ori harness reset droid. Reset removes only the entries, helper, and receipt that Ori wrote.- No batch models.
ori harness setup droidandori droidno longer add:batchmodels tocustomModels. Run setup again to remove the batch models that an earlier setup added.
Claude Code and Codex versions
- Claude Code version range.
ori claudeaccepts Claude Code from 2.1.116 up to, but not including, 3. A version outside that range exits 1 withORI_CLI_HARNESS_VERSION_UNSUPPORTEDand a hint. The hint says to runclaude updatefor an old version, orori updatefor 3.x and newer. - Claude Code auto-updater.
ori claudenow turns off the Claude Code auto-updater by default. To keep the auto-updater on, setDISABLE_AUTOUPDATERto an empty value, for exampleDISABLE_AUTOUPDATER= ori claude. Do not set it to0. - Claude Code tool search.
ori claudenow turns off tool search by default for each run. A shell or Claude settings value still wins, for exampleENABLE_TOOL_SEARCH=true ori claude. On the first launch after the upgrade, Ori removes the tool-search entries that earlier releases wrote to~/.claude.json. - Codex minimum version.
ori codexrefuses Codex below 0.118.0 withORI_CLI_HARNESS_VERSION_UNSUPPORTED. Older Codex ignores the providerauth.command, so every request fails with 401. The hint line prints thenpm install -g @openai/codex@latestandbrew upgrade --cask codexupgrade commands. - The checked binary is the launched binary.
ori claudeandori codexcheck the version of the binary that they launch, including one in~/.local/bin. On Windows,ori claudealso checksclaude.cmdunder a path that contains spaces.
Codex model catalog
- The
/modelpicker follows your guardrails. Inori codex, the/modelpicker lists only the tool-capable models that the key’s organization and API-key guardrails allow. - Separate catalogs for each launch. An API-key launch writes its own catalog file, so a launch under another key no longer overwrites the catalog that Codex Desktop reads.
- Codex Desktop matches its key.
ori harness setup codexwrites the Codex Desktop catalog from the models that the savedori logincredential allows. The catalog refresh uses the saved login. When no saved login resolves, the refresh fetches nothing and tries again in one hour.
Pi
- No
--modelon Pi 1.0.ori piworks without--modelon Pi 1.0.0. It starts on your saved Pi default when that default is an OpenRouter model, and otherwise on the OpenRouter default of Pi. - Offline mode.
ori pi --offlineandPI_OFFLINE=1 ori pimake no catalog requests and list the built-in OpenRouter models of Pi.ori pireadsPI_OFFLINEthe same way that Pi does, so1,true, andyesmean offline. - Faster start on a slow network. When OpenRouter is unreachable or hangs,
ori pireaches the prompt in about 3 seconds instead of 9 to 12 seconds. - Your models on hosts without IPv6. On a host where IPv6 has no route,
ori piandori pi --list-modelsnow load the models of your key instead of the built-in Pi list. - Built-in list notice. When the models of your key and the public catalog both fail to load at startup,
ori piprintsori: could not load your key's OpenRouter models; Pi is showing its built-in OpenRouter list.on stderr. Offline mode prints nothing. - Token refresh on 401. With
ORI_HARNESS_CREDENTIAL_HELPERon Pi 1.0,ori piretries an OpenRouter 401 one time with a new token. - Startup budget. On a slow machine, the time that Pi takes to boot no longer counts against
ORI_PI_STARTUP_CATALOG_BUDGET_MS.ori pilists the models of your key instead of the built-in Pi list. - Full model lists in a pipe.
ori pi --list-modelsprints every row when its output goes towc -l,grep,head, or another slow reader. - Pi subcommands.
ori pi mcp list,ori pi list,ori pi update,ori pi install <source>,ori pi remove <source>(oruninstall),ori pi config, andori pi authnow behave like the samepicommand. Ori’s--modeland--reasoning-effortdo not apply to them. - Rejected keys name
ori login. When OpenRouter rejects the key,ori piprints the 401 response and thenOpenRouter rejected this API key. Run `ori login` to sign in again., in-pprint mode and in the interactive error line. ori harness setup pi. Setup installs the Ori extension for Pi. It sets OpenRouter as the default provider andmoonshotai/kimi-k2.6as the default model, and keeps every other setting. Plainpithen uses your savedori login, also in a new agent folder thatPI_CODING_AGENT_DIRnames. Your own providers stay usable.- A saved login for Pi setup. Setup needs a saved login. An environment-only key fails with
ORI_CLI_OIDC_LOGIN_REQUIRED. ori harness reset pi. Reset removes the extension and the folders that setup created. It restores the originalsettings.jsonbytes.- Routing variants.
ori pi --model "anthropic/claude-haiku-4.5:online"and other routing variants now answer instead of failing with 400. A routing variant also works as the saved default model. The/modelcommand inside a session does not support variants yet. - Offline means no requests.
ori pi --offlineandPI_OFFLINE=1 ori pialso skip the launch settings download. A cached launch settings policy still applies.
Profiles
ori profile. A profile keeps the harness sessions of one context, such as work, apart from the others. In a terminal,ori profileopens a menu to switch, create, or remove profiles. Without a terminal, it prints the list.- Scriptable forms.
ori profile list,ori profile use <name>,ori profile add <name>, andori profile remove <name>do the same work without the menu. - Sessions location.
ori profile add <name> --sessionstakesshared, a template such as~/.{harness}-{profile}, or a path. The built-indefaultprofile keeps the shared sessions, so nothing changes until you add a profile. - Profile for one launch.
ori claude --profile <name>uses that profile for one run and leaves the active profile alone. A harness that cannot move its sessions folder keeps shared sessions and prints one notice. - Harness defaults.
ori profile use <name> --harness <kind>pins a profile to one harness.ori profile use default --harness <kind>clears the pin. A launch uses--profilefirst, then the harness default, then the active profile.
Sign-in and interns
-
Readable
ori authoutput.ori authnow prints its status as aligned fields, such assource,location,region, andendpoint, instead of sentences. The block is valid YAML.ori auth --jsonis unchanged. -
OIDC sign-in asks for intern access.
ori login --oidcnow requests access to manage your interns. The consent screen lists “Manage your interns”. -
Existing OIDC sign-ins sign in again. A session from before this release keeps working for inference in every harness. When
ori atorori interns attachfinds that the session cannot reach interns, Ori asks you to sign in again withori login --oidc. Accept, and Ori runs the login and lists your interns again. -
Errors without a terminal. When you decline or no terminal is attached, the command fails with
ORI_CLI_OIDC_LOGIN_REQUIREDand the hintRun `ori login --oidc` again to grant intern access.When no interns are visible after sign-in, the error isORI_CLI_INTERN_NONE_VISIBLE. -
Rejected API keys. A saved API key that the intern API refuses with 401 or 403 now fails with
ORI_CLI_INTERN_API_KEY_REJECTED. -
ori loginsaves globally.ori loginnow saves the login globally by default. Use--localto save it for the workspace only. Ori then adds a.gitignoreto the.orifolder of the workspace. The success line names the scope. -
Stale login recovery. When the saved login selects a session that no longer exists, Ori asks
Sign in again to replace it?in a terminal. Accept, and the command continues with the new login. -
A slow clock. OIDC sign-in no longer fails with
ORI_CLI_OIDC_RESPONSE_INVALIDwhen the computer clock runs behind. -
Which key wins. After
ori login, Ori says when an exportedOPENROUTER_API_KEYor another credential outranks the new login, and how to switch.ori atshows the key source when the key comes from the environment or a project dotenv file. -
Doctor finds the export.
ori doctornames the shell startup file and line that exportOPENROUTER_API_KEY, for example~/.zshrc line 12. Doctor also followssourcelines to files under your home folder. The fix offers to delete the line. -
Intern lookup errors. When
ori atfinds no intern with the name, it lists the visible interns one per line. The hint explains what to do when the intern is in another workspace. -
Unrecognized intern answers. When an intern answers with something that is not runtime info,
ori atfails withORI_CLI_DEV_RUNTIME_RESPONSE_UNRECOGNIZED. The error never prints the body of the answer. -
Personal connections in
ori at. Afterori login --oidc, remote terminal turns use your personal dashboard connections. An older OIDC grant withoutvault:readgets a 403 that tells you to runori login --oidcagain.
Administrator controls
- Allowed harnesses.
ORI_ENABLED_HARNESSESlists the launchers that stay available. Ori hides every other launcher, and a launch fails withORI_CLI_MANAGED_HARNESS_DISABLED.ori harness reset <kind>still runs, so a user can undo an earlier setup.ori doctorwarns about unknown entries. - Key limit ceiling.
ORI_KEY_LIMITsets the highest credit limit, in USD, for a key that Ori spends. Ori refuses a key with no limit or a higher limit withORI_CLI_MANAGED_AUTH_KEY_LIMIT_EXCEEDED.ori login --oidcasks for a key with exactly that limit, and the consent page shows it read-only. - Key limit reset.
ORI_KEY_LIMIT_RESETpins the reset period of the limit todaily,weekly, ormonthly. - Hidden consent fields.
ORI_HIDDEN_FIELDShideskeyLabel,expiration,creditLimit, orexistingKeyon the consent page. WithcreditLimit,ori auth,ori doctor, and refusals do not show the limit amount. - Workspace settings from OpenRouter. Before
ori <harness>andori harness setup <harness>, and afterori login, Ori downloads the launch settings of the workspace. Downloaded values rank directly below/etc/ori/config.json. A failed download never blocks a launch or a sign-in. A machine without a served policy behaves as before. - Regional key checks. With a saved region or
ORI_REQUIRED_REGION,ori login,ori auth region, and managed policy checks send the key check to the host of that region. - Custom API base.
ori login --with-keychecks the key againstORI_OPENROUTER_BASE_URLwhen that value is not an openrouter.ai host. - Rejected keys under a managed policy. A revoked or invalid
OPENROUTER_API_KEYunder a managed policy now fails withORI_CLI_OPENROUTER_KEY_REJECTEDand a hint to replace the key. Before, Ori reportedORI_CLI_MANAGED_AUTH_IDENTITY_UNAVAILABLE.
Interns and Slack
- Slack thread commands. Send
!q <message>to queue a message, or!steer <message>to steer the current run.!statusposts the model, the run state, and the queue.!helplists the commands. The model never sees these commands. - Mute a thread.
!mutemakes the intern answer only mentions in the thread, and!unmuteundoes it.!stopworks likestop. - Signed member posts. A message that the intern posts as a member ends with a line that names the intern. The Slack dashboard of the intern has a toggle for this line, which is on by default.
- Better Slack answers. The intern knows its own Slack name. It uses the questions form only when the form fits the question. It can explain where its files live, what a lost session lost, and when it answers in a channel.
- Continue a session in Slack.
POST /slack/session/threadopens a Slack thread on an existing session, and the next reply continues that session. In a direct message, the next top-level message also continues the session for 30 minutes. - Web page artifacts.
ori artifact save <file> --name <name>saves a self-contained web page in the workspace. On a hosted intern, the command prints the dashboard link to the page.ori artifact listandori artifact delete <name>manage the pages. - The intern reviews its own design. Ask the intern to “review how you’re built”, and the
architectskill runs.ori operations usagereports operations, model calls, and tool calls by feature and skill. - Trial copies.
ori trial BRANCHboots a test copy of the workspace on a branch with schedules off.ori trial sendsends the copy one message, andori trial stopremoves it. - Tools behind the egress proxy. Interns ship
curl.curl,pip,uv, and Python HTTPS calls trust the sidecar certificate throughSSL_CERT_FILEandREQUESTS_CA_BUNDLE. - Large backups. The
ori backupcommand of a hosted intern can push a pack of any size. Uploads over 10 MiB through the vault tunnel now go through. - Memory limits. An intern with a memory limit no longer refuses turns because of cached files.
CLI and configuration
- Region validation. Ori refuses an unusable
ORI_OPENROUTER_REGIONorregionconfig value before it sends any request. The accepted values areglobal,us,eu,usa, andeurope. The error code isORI_CLI_OPENROUTER_REGION_INVALID.ori auth regionandori doctorstill run, so you can repair the value. - Local daemon management.
ori daemon list,ori daemon status <id>, andori daemon stop <id>show and stop local daemons.ori daemon stop <id> --forcealso interrupts active runs. All commands support--json. - Region changes while running. When the region of a running
ori startbecomes unusable, the next run fails withORI_CLI_OPENROUTER_REGION_INVALIDand sends nothing. Fix the value, and the next run works without a restart.ori doctornames the source of the bad region. - Daemon telemetry on managed machines. On machines with a managed organization or workspace policy, or an unusable region, daemon telemetry no longer sends your OpenRouter key.
- Daemon shutdown finishes.
ori daemon stop,POST /api/daemon/shutdown, andori start --auto-update-restart exitnow end the process once teardown finishes, with the exit code of the request. If teardown hangs, the process exits 30 seconds after the request. - Daemon boot failures. When the daemon cannot start, for example because a feature cannot boot,
ori dev,ori start, andori evalexit withORI_CLI_DEV_RUNTIME_START_FAILEDand the cause. - Queued runs on one workspace. Runs that wait behind other runs on the same workspace now start instead of failing with
ORI_RUNTIME_PRE_HARNESS_TIMEOUT. - New feature folders in
ori dev.ori devno longer loads a feature folder that appears while it runs, and prints a warning. Restartori devto load the folder, or start it with--admit-new-featuresto load new folders on save. - Exit codes at startup. A Ctrl-C during startup now exits with code 130, and a SIGTERM exits with code 143. Before, the process ended by signal, which some wrappers report differently.
- Stable version without the commit hash. On a stable build,
ori versionandori --versionprint the version without the+<commit>suffix, for exampleori 0.15.7. Alpha builds still show the hash, andori version --jsonstill reports the full build version. - One shared code daemon at a time. When another
ori code daemon serveis still starting the shared code daemon after 30 seconds, a newori code daemon serveexits withORI_CLI_CODE_DAEMON_REFUSEDinstead of starting a second one. ori sessionslists real sessions.ori sessionsreads the current workspace, or the global workspace outside one, newest first in/resumeorder. It also lists theori codesessions that ran in the workspace root.- New
ori sessionsJSON. The--jsonentries now use the/resumesummary fields, such aslatestPrompt,startedAt,endedAt, andusage.costUsd. - Cleaner command output. A bare group command, such as
ori workspaceorori daemon, prints its help. A usage error prints one line and a hint. Ctrl-C at a prompt printsCancelled.and exits 130. - Version and panel output.
ori versionprintsori <version>.ori logs --panelnow needs a terminal and fails withORI_CLI_TERMINAL_REQUIREDotherwise. - Dotenv keys in
ori code. In a repository whose.envsetsOPENROUTER_API_KEY,ori codestarts the shared daemon after you confirm the key one time. ori devexits cleanly. Quittingori devalso stops itsori tuiandori logschildren.ori starton Windows.ori startno longer crashes at startup on Windows.- Log levels. Failures in feature code log at WARN with codes such as
ORI_SCHEDULE_USER_CODE_FAILED,ORI_TOOL_CALL_FAILED, andORI_FEATURE_STATE_QUERY_FAILED. ERROR now means that Ori itself failed. - Console logs with OTLP. With
OTEL_EXPORTER_OTLP_ENDPOINTset, the console prints one line for each record at WARN and above.
Agent runs
- Steering between tool calls. A steer that arrives during a tool batch takes effect after the current tool, and the rest of the batch does not run. Queued messages still wait for the end of the turn.
- Stalled commentary stops. When the decisions model judges that repeated commentary-only responses have stalled, the turn stops as
cancelled. Ori keeps the history, so “continue” resumes the work. - Shell output on failure. A shell command that times out keeps its partial output. Large output keeps the start and the end.
- Signals and long output. A command that a signal kills returns its output with
[killed by signal <NAME>]. Streamed output over 30,000 characters ends with[live output truncated at 30000 characters]. - Shell commands yield the CPU. The
shelltool runs commands at the lowest CPU priority. A timed-out command that is still alive 5 seconds after SIGTERM gets SIGKILL. - Large skill folders. Turns with large or slow skill folders now start instead of failing with
ORI_RUNTIME_PRE_HARNESS_TIMEOUT.
Models and failures
- Allowed-providers failure. When no allowed provider serves the selected model, Ori reports
ORI_MODEL_GUARDRAIL_RESTRICTEDinstead ofORI_MODEL_UNAVAILABLE. - Daily free-model limit. When the account reaches its daily free-model requests, Ori reports
ORI_OPENROUTER_FREE_LIMIT_REACHED. Ori then switches to the fallback models at once instead of retrying the free model. - No false model warning at launch.
ori <agent> --model <id>now checks the model against the models your key can use, not only the public catalog. A model that only your key can use, for exampleori pi --model cohere/north-mini-code:free, starts with no catalog notice. A model that neither catalog lists still passes through with the same notice. - Guardrail 404. When the guardrail or data policy of the workspace excludes every endpoint for the model, the run fails with
ORI_MODEL_GUARDRAIL_RESTRICTEDinstead ofORI_MODEL_UNAVAILABLE. Ori then tries the configured fallback models.
Daemon and feature reload
- Rejected reloads are visible.
GET /api/featuresreports"status": "previous_version"for a feature whose latest edit did not load. A feature that a rejected reload added shows asneeds_repair. The next successful reload clears both statuses. - The intern sees features that need repair. On its next turn, the intern prompt lists each feature whose last edit did not load.
- Model changes reach running chats. After a model change on a running
ori start --watchintern, the Slack agent reports the new model on its next turn.
Schedules and usage
- Usage by feature and schedule.
GET /api/operations/usagereturns operations, failures, runs, tokens, and cost for each feature, operation name, and model by UTC day. Each schedule fire counts as one operation, and a run that a hook starts counts ashook. - Cost for each feature. Model calls are now billed to the feature that made them. This covers direct OpenRouter calls from schedules, hooks, routes, and tools, chat turns, and turns that feature code posts to
/api/invoke.GET /api/operations/usagealso returns acallslist with the model, tokens, and cost of each call, by trigger, feature, and skill. - Models for schedules and features. A schedule takes
modelin its definition or Markdown frontmatter.PUT /api/schedules/:name/modelsets the model of one schedule, andPUT /api/features/:featureId/modelsets the model of a feature. - Pause or resume a code schedule.
PATCH /api/schedules/:namewithdisablednow works for schedules defined in code. A stored"disabled": falseturns on a schedule that its definition disables. - No retries on a refused account. When a schedule run ends in a failure that is not retryable, such as
ORI_OPENROUTER_CREDITS_EXHAUSTED, Ori records one failed run. The schedule waits for its next cron time.
Evals
- Crashed runs count as errored.
ori evalreports a run that never answered as errored, never as a pass. A model whose every run errored readsunmeasured.ori eval --jsonrows gain aresultfield. - Judge failures leave runs ungraded. When the judge call fails, the run reads
ungradedwith the judge error, for example3/4 passed (1 ungraded). An empty answer now fails. Runori eval scratchagain to get this behavior in a scratch workspace. - A clear judge error. A failed judge call names its own failure code instead of an unparseable verdict.
- Careful conclusions. The conclusion declines to pick a model when more than a tenth of its runs errored, went unmeasured, or stayed ungraded.
- Wide panels. Cases past 64 concurrent runs now wait for a free slot instead of failing with
ORI_HISTORY_CAPACITY.
Feature authoring
- Feature webhooks. A feature declares webhooks in its
apicontribution with aschemeand a vaultsecretname. Ori serves each webhook at/__ori/hooks/<feature>/<name>. The handler runs only after the vault verifies the delivery. - Webhook details.
GET /api/featureslists the webhooks of each feature.ctx.webhook.urlgives the handler its public URL. - Retry signals. A temporary verifier problem answers
503or429withRetry-After: 5. A401always means a wrong signature, timestamp, or secret. Non-Slack webhooks wait up to 5 seconds for verification. - Slack events webhook.
POST https://<intern>/__ori/hooks/slack/eventsserves Slack events behind vault verification. - Session routes.
GET /api/sessions/:id/conversationreturns the conversation of a session without every event.GET /api/sessions/persisted?view=summaryreturns a summary listing. - Session source. Each run records its surface and user.
context.callerAuthorizedtells a route whether the request carried the daemon credential. - SDK schemas. The
oriauthor SDK exports more schemas, such asAgentInputSchemaandThreadContextSchema. Every existing export keeps its name.
Chat TUI and MCP
/resumeduring a turn. Inori code,/resumeand feature slash commands run while the agent is still working./compactwaits until the current turn ends.- Version in the banner. The
ori codebanner shows the Ori version, for exampleOri Code v0.15.7. - Terminal.app colors. On macOS Terminal.app,
ori code, theori logspanel, and the schedules run view use 24-bit color only whenCOLORTERM=truecoloris set. Otherwise they use 256 colors. - Light theme detection. When no theme is set,
ori codeuses OR Light on a light terminal profile. - MCP optional placeholders. When the model sends an empty placeholder for an optional MCP argument, the retry omits that argument and the call succeeds.
- MCP sign-in page. After you select Connect in
/mcp, the browser shows the same page style asori login, with light and dark themes. The failure page tells you to retry Connect in the MCP panel.
Release assets
- Source archive. Each CLI release now publishes an
ori-source.tar.gzasset, whichSHA256SUMSlists. The archive extracts, installs withbun install --frozen-lockfile, and builds the CLI withbun run compile:cli.
This release adds Remote Control for interns through
ori interns and ori at. OpenRouter OIDC sign-in becomes a direct credential in every harness. ori codex and ori claude gain cloud session import, cross-provider resume, and a relay that keeps Codex on OpenRouter. Schedules get an HTTP API, Slack delivery, and an acting member. The release also adds new launchers, a private terminal drawer in the TUI, ori migrate, and many fixes for daemon memory, Windows, and failure reporting.Remote Control and interns
-
Personal connections in the terminal. After
ori login --oidc,ori atsigns terminal turns as the logged-in member, enabling personal dashboard connections. API-key sessions keep shared access and explain how to enable personal access. Requires the matching intern API deployment. -
ori internsis the canonical command. In a terminal,ori internsopens a picker and attaches to the intern you choose.ori interns listprints one intern per line and never prompts.ori interns attach <intern> [--resume <id>]attaches directly.ori atstays as an alias, and the word afterinternsis always a subcommand. -
Open an intern by name or slug.
ori ataccepts an intern slug, and the picker filters by name as you type. An empty intern list offers login and shows the scope of the list. A rate-limited request tells you to wait. -
Remote TUI connections are safer. A bare remote host defaults to HTTPS, and the runtime probe has a time bound. The CLI refuses a non-loopback
http://target whenORI_DAEMON_TOKENis set. Outage HTML from the server no longer appears in the remote TUI. - Remote sessions survive restarts. A remote TUI stays open when its intern daemon restarts mid-turn. When the intern goes silent, the TUI stops with a clear message.
- Attached sessions are marked. The chat TUI shows a clickable intern tag on an attached session, and the tag runs to the right edge.
Sign-in and credentials
- OIDC is a direct credential. A saved OIDC session from
ori login --oidcworks like a saved API key. This covers every launcher, includingori claude,ori codex, andori pi. It also coversori code, the daemon,ori doctor, andori auth. Ori no longer runs a separate broker for OIDC. - Tokens refresh live. Ori refreshes the access token before it expires, in every harness. Claude Code, Claude Desktop, and Codex read a fresh token through one Ori credential helper, so long sessions keep working. A rejected refresh fails with
ORI_CLI_OIDC_LOGIN_REQUIREDand namesori login --oidc. - OIDC edge cases. Ori refuses an OIDC session that an administrator disabled. A 401 during model or preset loading refreshes the saved token. When OpenRouter refuses an OIDC session for the model catalog, Ori reads the public catalog.
- Connect sign-in is clearer. The
ori loginoutput for Connect sign-in is short and easy to scan. The paste prompt clears and stray input is discarded once sign-in succeeds. Ori warns when a higher-precedence credential overrides a new Connect sign-in. Pasted login codes are checked before use. - Regional sign-in uses the global auth page.
- Managed policy controls. A managed policy can lock Ori to one OpenRouter region and can choose the credential storage custody. Ori refuses OIDC broker leases once a managed policy is installed. Under a managed policy, Ori verifies the owner of the OpenRouter key for interactive CLI use. On Windows, Ori reads the managed policy from a fixed ProgramData path.
- Key admission is stricter. Ori admits the OpenRouter key before
ori doctor,ori auth region,ori auth, andori loginuse it. Ori refuses to send an OpenRouter key to an endpoint or transport that a project dotenv chose. The saved key stays out of child processes, and Ori no longer trusts the owner cache file. - Shared credential unlock. Login and MCP share one user-presence protection. Protected commands,
ori interns, andori athold the shared unlock while they run. - Linux storage without a Secret Service. On Linux without a Secret Service, Ori keeps device keys in systemd user credentials. This path works without
busctland reports typedsystemd-credsfailures. - Credential storage is more robust. Keychain reads time out after 10 seconds headless and 30 seconds at a terminal instead of hanging. Ori names an absent OS credential service and checks it before sign-in. Reset revokes the credential and clears dangling selections. Credential files are written atomically, and late key store corruption is quarantined. A dangling saved login reports as signed out. The credential-source line shows only for a project
.env. - The login callback is hardened. The loopback callback and the browser opener resist local attackers.
- MCP passphrase fallback. MCP credential storage has an explicit passphrase fallback and storage diagnostics.
Codex
ori harness setup codexreplacesori codex-desktop. The command configures Codex and Codex Desktop to use OpenRouter, and it saves your previous setup socodex --profile originalstill runs it.ori harness remove codexundoes the setup. Setup works with OIDC, follows your region, and writes a hosted Codex model catalog.- Setup is safer. Setup finishes when Codex holds its thread index locked, and it refuses a read-only
config.toml.ori doctorreports a missing helper path, and--fixrewrites it. After a reset, Ori says that the setup is still active and lists the lines to delete. - The relay keeps Codex on OpenRouter.
ori codexkeeps theopenaiprovider ID through a local relay and serves web search through OpenRouter. The relay hides the ChatGPT plan quota, stops retrying a rejected OpenRouter key, and shows OpenRouter’s own reason after a 402 or 403. - Resume across providers.
ori codex resumelists threads from all providers, andori codex exec resume --lastresolves across providers. Resume works after Codex options and waits for the Codex background server. - Codex Cloud tasks. Ori imports Codex Cloud tasks into local Codex sessions.
- The model picker lists OpenRouter models.
- Provider and config fixes.
ori codexuses its own provider ID, so a useropenroutertable cannot break it. The provider stays in place when-cfollows a subcommand. Codex TOML strings escape control characters, and-coverride values keep an apostrophe. The websocket fallback error is quiet inori codex execandori codex review. Codex launches without daemon auto-start, so the embedded-mode warning is gone.
Claude Code and Claude Desktop
- Accounts and profiles.
ori claude --profilepicks a Claude account folder.ori claudeshows sessions from other Claude accounts and copies MCP logins into a per-account store. - Cloud sessions.
ori clauderesumes claude.ai cloud sessions, and imported cloud sessions appear in the/resumepicker. Ori stops cleanly when a cloud session cannot resume. - Settings are respected.
ori claudereads Claude settings fromCLAUDE_CONFIG_DIR, merges a caller’s--settingsinto its injected settings, and keeps your Claude profile on Connect launches. Claude Code keeps its native prompt and tool surface. - Model handling. Claude Code gets its own model ID for OpenRouter
anthropic/*slugs, and it no longer remaps legacy model IDs on every launch. Tool search stays on for Claude and off for every other model. Ori blanksx-api-key, so OpenRouter no longer pins Claude to Anthropic. When the catalog fetch fails, Ori reuses the last known model catalog. - First launch is quieter. Claude no longer asks about the OpenRouter key on first launch, and Ori no longer warns about MCP login stubs that hold no token.
- Claude Desktop.
ori claude-desktopregisters on a Connect lease and accepts a version range with launch contract checks. Ori detects Claude Desktop updates and records the version at launch.
Other launchers and harnesses
- New launchers.
ori dshlaunches the DeepSeek Harness server.ori crushlaunches the crush harness. Grok, Hermes, unreal, and zcode launch through external adapters. - Carry a setup across harnesses.
ori harness migrate --from claude --to codexprints a dry-run report of how instructions, skills, and MCP servers move to Codex.--applywrites the migration, and--scopepicks the project or user setup. MCP secret values are never written. - Route plain commands through Ori.
ori alias --binarywrites a script for each selected harness, so a plainclaudeorcodexcommand runs through Ori. The real binaries move to.ori-real/<name>, and rollback is safer. ori clineshows your own Cline sessions, history, and MCP servers.- OpenCode and Kilo. OpenCode 2 launches with the model in config and a private server.
--reasoning-effortreaches OpenCode 2runandminias-m model#variant. Everyori opencodeandori kilorequest is attributed to its Ori session. - Regional routing is unified. Every harness launch follows the saved OpenRouter region, including cline, prime-agent, codex, and Hermes. Opencode, cline, and omp use the verified OpenRouter endpoint. Every Anthropic model that
ori piroutes through OpenRouter has a pinnedbaseUrl. ori pimodels. The OpenRouter catalog replaces the built-in OpenRouter models inori piat startup.- Missing tools are named.
ori dshreports a missing pnpm before it runs plugin commands.
Chat TUI
- Private terminal drawer.
/terminal <command>or the agent’scommand-terminalrequest opens a terminal drawer at the bottom of the TUI. Press Ctrl+] to open or leave the drawer, and press Enter to launch the command. Input and output stay in the TUI, and the agent receives only the exit result. The drawer works on macOS and Linux. - Click a file path to open that file.
- Resume and usage. The TUI resumes the active session and shows the resolved usage model. Resume previews appear in the main chat.
- Clearer tool and approval rows. A tool call that an approval refused shows as Rejected. The permission decision line appears when you answer the card. A failed steer send names the interrupt fallback.
CLI
ori sessionslists recorded sessions.ori sessionsnow does whatori historydid, with--archived,--limit, and--search.ori sessions --runninglists the sessions a live runtime holds. Scripts that ran a bareori sessionsfor the runtime list now need--running.ori historyremains a hidden alias.ori sessionsoutput. When piped,ori sessionsprints a JSON envelope, and it fails on a bad store.ori doctorruns every check when you run it with no arguments.ori version --semverprints only the version number.- Minimum supported version. A CLI below the minimum supported version prints a message that tells you to upgrade.
ori migraterewrites feature shapes that a release removed. It runs each rewritten feature’s tests and reverts a feature whose tests fail.--dry-runprints the plan.- Prompts and errors. Ctrl-C at an
oriprompt exits quietly with code 130. An invalid global flag value reports as a usage error. Config warnings stay off stdout.
Schedules and daemon
- Schedules over HTTP. The daemon creates, edits, toggles, deletes, and triggers schedules over HTTP. It also lists pending runs, cancels runs, and sends results to Slack.
- Schedules act as a member. A scheduled run acts as the member who saved it, and the acting member survives later saves. The schedule reports its approval state.
- Scoped MCP servers. A schedule can limit its turns to chosen MCP servers.
- Feature information.
GET /api/featuresreports description, source, skills, schedules, and feature status. The daemon serves a feature’s files and skill text read-only, and runtime info reports the declaring workspace. - Runs and sessions. The daemon closes runs that an earlier daemon left open. It exports the bound port as
ORI_RUNTIME_PORTinto every run. You can resume another session by its short ID. - Memory and shutdown. Each intern runs on one lean
oriprocess. The daemon no longer keeps every session it has served, and it frees the previous feature build on every reload. It reads one session’s events without loading every log. On SIGTERM, the runtime drains turns, and it admits turns by memory. - Faster session history. Sessions are admitted in constant time, and history cleanup runs on a worker. The retention sweep walks history once per sweep. A refused session history store names its mismatch and backs up under the lock.
- Shared Code daemon.
ori coderuns one shared Code daemon per user with detachable sessions that the terminal owns. The daemon stays alive across laptop sleep and routes each run by its caller’s region. - Previews. Each intern gets warm localhost previews and a headless Chromium.
- Skill cache fallback. When
TMPDIRrejects embedded skills, the runloop falls back to a home cache.
Slack
- Auto-responders on the dashboard. The dashboard shows auto-responders.
- Lookups as the member who asked. On request, a lookup runs as the member who asked, goes through the vault, and lists that member’s conversations. Runs that Slack did not start also get the lookup skill.
- HTTP MCP as the member. HTTP MCP connections open as the member whose turn it is.
- Threads and messages. A thread that the agent spawns is engaged, so unaddressed replies reach the agent. A message whose first word is
asideis ignored. Basic answers keep the intern’s internals out. - Safer resume. An interrupted turn resumes at most once, so a crashing turn cannot loop.
- Skill scripts use the egress proxy for Slack calls.
Feature authoring
- Declared dependencies. A declared dependency is the only way one feature reaches another. A misdeclared feature dependency surfaces at boot.
- Feature checks and pricing.
ori features gateproves that a feature set fits the framework. The runtime prices every operation a feature names. Feature routes declare per-route access. - Workspace packages. Installed workspace feature packages boot without a registry. An added feature loads without a restart, and interns learn its new commands.
- Vault preflight. Interns run a vault preflight for features.
- Upgrade stability. Intern features keep loading across Effect upgrades.
Native harness
- Context overflow compacts. A provider context-overflow error code triggers compaction.
- Failed tool work fails the turn. A turn that ends with unanswered failed tool work fails. A routed short-prompt command that fails also fails the turn.
- Model names. A switch-model answer names the model that
openrouter/autorouted to. - Vault credential reasons. Ori names why the vault left an MCP server without its credential.
- Server tools by region. Ori offers OpenRouter server tools only in regions that serve them.
- Resume and compact. Resume cost is guarded, and
/compactno longer seeds an empty child. - Skills with YAML indicators. Ported skills load when one-line frontmatter values hold YAML indicators. Malformed YAML frontmatter reports as one diagnostic.
Telemetry and failure reporting
- Failure causes. Every failed
orirun names its cause, model, and Ori session.agent_runandeval_runreport the failure cause. A failing eval test counts as quality, and every infra cause is named. A failed schedule fire tells Datadog which schedule failed and why. - More signal.
tool.failedlog records carry the typed tool failure code. Telemetry shows where harness setup is abandoned and which prompts a human typed. Ori forwards all telemetry from every wrapped CLI that supports OTel. TUI usage events come from the CLI process. - Vault tunnel outcomes. Every sidecar exchange records exactly one outcome, and vault-tunnel configuration errors count as user failures. ALPN uses the wire format for Bun’s TLSSocket.
- Dev runtime errors. Ori says why it could not reach its dev runtime.
Backup
- Backup no longer includes
state.sqlite.ori backuprefuses to run mid-merge and commits swapped restores. - Clear failures. Every failed backup step says what git said, and each failure has its own code. Ori reports a missing backup remote and warns once when the remote refuses authentication.
Windows and platforms
- Windows. Features and npm harnesses boot on Windows, and Ori children and bun installs run from the compiled binary. Template archives unpack in-process, so Windows scaffolds skip links. Windows host paths resolve correctly.
- Musl and gVisor. Musl binaries select the musl OpenTUI native package. OpenTUI output streams through stdout under gVisor.
- Child spawns. A child spawn that the OS refuses fails instead of crashing Ori.
Release channel
- Stable advances safely. The stable pointer advances only when it contains every published stable.
Sign-in moves onto one encrypted credential engine, with OpenRouter Connect device and browser flows, encrypted saved logins, and a biometric local vault. The native harness defaults to
anthropic/claude-opus-5.5 and recovers rejected, oversized, and guardrailed requests with model fallbacks chosen by Jev. Slack answers unaddressed messages through an auto-respond pipeline, resumes interrupted turns after a restart, and opens turn details in a modal. The daemon keeps detached runs alive, reloads at a run boundary, and answers /health before features finish booting.Sign-in and credentials
- OpenRouter Connect sign-in.
ori loginsigns in through OpenRouter Connect with a device code on a machine without a browser, or with an authorization code and PKCE where a browser is available. The broker that completes the exchange runs on the shared credential engine. - Saved logins are encrypted. A saved login is stored as an encrypted record instead of a plaintext key. The same record engine now serves the MCP credential store and the vault, so one encryption path covers all three. Run
ori auth migrate --localorori auth migrate --globalto convert an existing saved login. A save that cannot reach the operating system credential store fails without writing a plaintext key. Pass--insecure-storageto save plaintext on purpose. - Biometric local vault. The vault can hold its key behind a local biometric prompt on a machine whose operating system supports it.
- Managed policy recovery. A stored credential that managed policy rejects prompts for reauthentication instead of failing the command. An external launcher stays available under managed authentication policy.
- Concurrent credential writes. Credential selection writes take an operating system lock, so two Ori processes no longer overwrite each other’s choice.
- MCP asks for credentials later. An MCP server’s credentials are read when a tool search needs them, not at startup. A user-presence custody provider can hold an MCP key behind the shared engine.
Slack
- Auto-respond pipeline. A feature registers an auto-responder through the
slack.autoRespondhook and theautoRespondregistry. Jev classifies an unaddressed message in a registered channel, and a positive verdict starts a turn. - Interrupted turns resume. A turn cut off by a runtime restart continues after the restart instead of going silent.
- Turn details in a modal. The See details button opens the turn details in a Slack modal.
- The thread hears about an unreachable model. Model retries are bounded, and the thread gets a message when the model stays unreachable. The unreachable clock starts only for model retries, so an unrelated wait no longer triggers that message.
- Signed request freshness. Ori refuses a signed Slack request dated more than five minutes in the future.
- Other bots are quoted correctly. A thread message from another bot is attributed to that agent rather than to the person who started the thread.
- Stored sessions with no history. A stored session that holds no history starts fresh instead of resuming an empty conversation.
- Requester identity reaches OpenRouter. The harness sends the Slack requester as the OpenRouter request user.
Native harness and models
- The default model is
anthropic/claude-opus-5.5. - Jev selects recovery fallbacks. A request the provider rejects, a context that exceeds the model window, and a guardrail refusal each recover on a bounded fallback that Jev picks. The recovery comparison runs without an opt-in.
- An unknown model ID fails. A model ID the catalog does not know reports a failure instead of silently falling back to another model.
- Cloudflare origin failures are named. An origin-unreachable status from Cloudflare is classified as
network_failed. - Provider detail on model failures. A model failure carries provider detail for telemetry.
- Tool arguments survive the call. Optional tool arguments are preserved. A rejected MCP call retries without the rejected nulls and undeclared keys.
- MCP tool names are prefixed. The model sees a prefixed name for each MCP tool, so two servers offering the same tool name stay distinct.
- First stdio MCP launch gets 90 seconds. A stdio MCP launcher has 90 seconds to answer on its first run, which covers a cold install.
Chat TUI
- The resume picker previews conversations. Each entry in the resume picker shows a preview of its conversation.
- Forked sessions keep turn controls. A forked session has the same turn controls as any other session.
Daemon and runtime
- Detached runs stay alive. A detached run continues after its caller disconnects.
- Reload waits for a run boundary. A reload happens at the next run boundary instead of killing a run in flight.
/healthanswers during boot. The health route responds before the first feature finishes booting.- Steering and interactions. A steer sent while the run waits on an interaction is refused at once. A response to an interaction that no paused run holds answers
unknown_interaction. - Shared workspaces are announced. An intern is told when its workspace is shared with another intern.
- The harness anchors to the workspace root. A composed feature set no longer moves the harness off the workspace root, and the skill snapshot stays beside the composed features.
- A feature module evaluates once per boot. Each feature module is evaluated one time per runtime boot.
- An agent-started Ori is not the intern. An Ori process the agent starts reports as itself instead of as the intern.
- A corrupt state store rebuilds. A corrupt SQLite state store is rebuilt once instead of failing every write.
- Failed schedule runs carry a cause. Every failed schedule run records an
error_class. The runtime logs the session-history database it resolved when it opens one.
CLI, launchers, and configuration
- ZCode launcher.
ori zcodelaunches ZCode as an official harness. - Unreal agent launcher. A launcher for the unreal-agent harness ships with the CLI.
ori.mdmodel fields.ori.mdresolves a variant slug and accepts provider request fields.- A drifted SDK is detected by version. The CLI compares the installed SDK version instead of checking that the dependency exists, so a stale SDK is reported.
- A bad feature include no longer stops boot. A declared features source with a bad include is skipped, and the remaining sources load.
Backup, sandbox, and feature authoring
- Backup needs a git repository.
ori backupskips its scheduled fire when the workspace is not a git repository.ori initpersists a fallback git identity so a backup can commit. The vault proxy git configuration is applied to every process Ori starts. - Protected paths stay unreadable. A sandboxed command cannot read a protected path through a symlink or a case variant of its name.
- Typed built-in hooks. A feature subscribes to a built-in hook with its declared types, and
use("slack")returns a typed Slack handle.
Session history stores open with indexes on every reverse reference, so history retention no longer scans whole tables and a large history no longer stalls a session read.
Session history
- Retention checks use indexes. Opening a new or existing history store installs ten indexes on the reverse references that the retention sweep and deletion guards consult, such as chunk children, checkpoint parents, session heads, and merge parents. Before this release, some of those checks scanned the whole table, and a large history could stall an agent reading its session history. The schema version stays at 4, an existing store needs no migration or backup, and retention semantics are unchanged. Retained session lineages can still grow, since this release adds no byte bound on history storage.
A slow SQLite call no longer freezes the daemon, and a stuck state store call fails after 30 seconds instead of hanging. The bundled MCP setup skill now describes the native OAuth flow instead of asking for an API key or calling OAuth unavailable.
Daemon
- State store calls no longer block the daemon. A slow or lock-contended SQLite call in the builtin state store runs off the main thread, so HTTP handling, timers, and runtime progress keep going while it waits. Each store call has one 30-second deadline that starts when the call is made and covers queue wait, directory creation, lazy open, and the SQL itself. Reads, writes, and close run in method-call order, including calls submitted from completion callbacks. Existing SQLite files, keys, tables, foreign keys, and WAL behavior are unchanged, and no migration is needed. A write that times out has an unknown outcome and is never replayed. The separate session history database is not affected.
Feature authoring
- The MCP setup skill explains OAuth. Ask
Add Linear globally using OAuthand the skill describes~/.ori/mcp.jsonwithauth: { "type": "oauth" }and browser consent through/mcp, instead of asking for an API key or claiming OAuth is unavailable. The guidance covers encrypted saved grants, workspace-bound credentials, matching daemon and chat working directories, and vault or proxy restrictions. It also tells the agent to check the current implementation or release documentation before it declares a capability unsupported.
The native Ori harness recovers from an oversized model request by redacting older images, reports an open model request on the event stream, and uses Jev decisions for tool search, skill suggestions, typed questions, and short TUI prompts when the catalog advertises them. The chat TUI marks queued prompts with a
[Q] chip and keeps its diagnostics after a bad boot. ori logs --follow carries lifecycle and failures instead of token narration, a crashing daemon still exports its run trace, and an unreachable or misconfigured OpenTelemetry collector is named in a warning. Slack turns act as their sender again on identity-layout interns, and a hosted intern without a backup remote no longer reports failed backups.Native harness
- Oversized requests retry without images. When the encoded model input exceeds the egress ceiling,
ori coderetries once with image attachments and tool-read images replaced by placeholders. Text attachments stay intact and no flag is needed. Later turns and forks start from the redacted checkpoint, so those images stay omitted after a failed retry or a restart. New attachments still work. Input that stays oversized without images fails before Ori sends it. - An open model request reports progress. A new
model.progressevent appears on the invoke stream, the session event stream, and the journal while a model request is open and nothing has come back. Its payload carrieselapsedMsanddeadlineMs, both reset per attempt. It stops at the first model event and emits at mostsilence / intervalevents per attempt, 24 at the defaults. The event joins the OTLP lifecycle records withori.model.elapsed_msandori.model.deadline_ms, and theori_runtime_events_total{type="model.progress"}counter. The chat TUI adds no transcript line for it, because its working indicator already shows elapsed seconds. - Jev decisions where the catalog advertises them. When the configured OpenRouter user catalog advertises Jev,
tool-searchranks MCP tools semantically, with keyword search as the fallback.skilloffers advisory suggestions without hiding eligible skills or loading their bodies.ask-user-questionuses typed choices in the existing unanswered-question fallback. Empty queries and exact tool names bypass inference. Manual mode, caller-owned interactions, and free-text questions keep their behavior. Availability comes from/models/user?output_modalities=decisionsand never from paid probes. A positive result lasts ten minutes, a negative result or failure one minute, and inference has a 1.5 second deadline. Regional settings such asORI_OPENROUTER_REGION=eustill apply and never fall back globally. The US and EU public catalogs currently advertise no decision models. - Recovery shadow comparisons, off by default. Set
ORI_JEV_RECOVERY_SHADOW=1in the environment or underenvin.ori/config.json. A root run then compares retry against configured model fallback after repeated transient failures, one comparison per logical request, three per root run, 1500 ms per call, no retries. Child threads do not compare. Diagnostic logs record the scores, the action taken, the outcome, and decision usage. Decision spend is separate from thread totals, and a missing cost is recorded as unknown. No recovery-quality improvement is claimed.
Chat TUI
- Short prompts can run a command. With the Ori harness and an available Jev route, a single-line text prompt of at most 200 characters can run a TUI command directly. Typing
show my token usagerequests/usage. Both decisions need a confidence of at least 0.9, and an unavailable, invalid, uncertain, or abstaining decision falls back to the usual model turn. Commands with arguments, including/model, are outside this first slice. An eligible prompt spends one paid decision request that contains both questions, and direct execution needs no chat-model request. - Queued prompts carry a
[Q]chip. The rows above the composer render[Q] <prompt>instead of an hourglass glyph. Unselected chips sit on the dim color, and the selected row is highlighted end to end in the accent color. The keys are unchanged. Arrows select,xdeletes, Enter steers, and Esc leaves. - Diagnostics survive a bad boot. When
~/.ori/logscould not be created at start, the next flush retries instead of writing nothing for the rest of the process. On exit,[ori] N chat TUI diagnostics could not be written to <path>prints wheneverN > 0, andORI_TUI_DIAGNOSTICS_SUMMARY=1still gates only the success summary. A thread event the runtime cannot decode is reported under theharnessscope with its own discriminant and the decode detail, so it reachesori logsandGET /api/logs/stream. A state-store failure carrieserror_codeanderror_errnobesideerror_name, with no key, value, or SQL text.
Logging
- The live log stream is readable.
ori logs --followandGET /api/logs/streamcarry the session’s lifecycle and failures. Streaming-token narration, which made up most of the hub lines on a measured four-turn session, no longer reaches them. ORI_LOG_LEVELis the one floor.ORI_LOG_LEVEL=debugis the documented way to putdebugrecords in front of the collector. Effect’s own--log-levelstill works, and the framework only lowers the floor it inherits.ori dev --splitkeeps its panes. A harness warning mid-turn goes to the events pane instead of over the chat pane.ori code --output jsonlstays pure JSONL. A strayconsole.logcannot corrupt the protocol stream, and ordinary diagnostics still reach the terminal.
OpenTelemetry
- A crashed daemon still exports its run. With
OTEL_EXPORTER_OTLP_ENDPOINTset, an uncaught exception still produces theori.runspan for the run in flight. The span is markedori.run.incompletewhen the daemon died mid-turn, and thedaemon.uncaught_exceptionFATAL record arrives as before. The newori_otlp_records_pushed{signal}metric counts the records Ori handed to an exporter buffer, including those a disabled exporter drops. - A broken collector says so. With an unreachable endpoint, each signal logs one WARN per 60 seconds naming
ori.otel.signal,ori.otel.export.reason, and the cause, such asECONNREFUSED,ENOTFOUND, an HTTP status, or the collector’s body. A collector rejection now surfaces the collector’s ownerrorMessage. A scheme-less value such asOTEL_EXPORTER_OTLP_ENDPOINT=localhost:4318keeps export off and warns once, naming the variable and the value’s length, never the value. An unset variable stays silent. ori doctor <harness>reports export configuration. A new check says whether export is disabled, unconfigured, unusable, or usable, reported by origin so credentials in the URL stay out of the output. It does not probe the collector.- A large accepted response is not read as full acceptance. Ori reads a bounded 8 KiB prefix of an accepted body. When a
partialSuccessresponse is larger than that, Ori warns that the collector accepted the batch and threw records away, withori.otel.export.detailsaying the rejected-record count was not seen. - The run-trace subscriber counts dropped entries.
ori_journal_entries_missed{subscriber:run_trace}counts the journal entries its bounded live tail dropped. - A cancelled turn is journaled. A cancel or disconnect during an invoke stream writes a
turn.failedwith failure codecancelledto the run journal and the lifecycle logs. It carries the resolved harness and the interrupted turn id, and therun stalledandrun evictedrecords gainori.run.last_event_type. - The vault names a response that ended early. The sidecar log reads
the vault response body ended after N byteswith the socket message and code, such asECONNRESET, instead of a bare decode error.
Usage telemetry
- A failed command reports its failure. A command that ends with a non-zero exit reports
cli_commandwithoutcome: "error"and emitscli_errorwithexit_code. Itsfailure_kindseparates a turn that ran and failed from Ori breaking. Every direct harness launch such asori claudeorori codexemits onecli_commandwith the command alone, and a process emits at most one.install_first_runreportsinstallerorsourceinstead ofunknown, andfeature_invokedcounts the feature id rather than the typed command name. cli_errorcarries no hostnames or workspace paths. A URL is reported as<scheme>://<host-class>/, where the host class isopenrouter,loopback, orother. A path outside any home directory is reduced to its basename. A bare hostname in a network failure message is reported as<host:other>,<host:loopback>, or<host:openrouter>. Home-directory paths still become~or<user-path>.- Exit telemetry is delivered before it is called sent. The daemon’s exit telemetry leaves through a detached child, as the CLI’s already did, so a daemon that ends the moment a turn finishes still delivers that turn’s
agent_run. A new~/.ori/telemetry-delivery/directory holds one small receipt per exiting process, with two booleans and the install id, never event content. The nextorirun folds the receipts intotelemetry.jsonand deletes them. An unreadable receipt is read again next boot, bounded at 16 files. ORI_TELEMETRY=0still disables all of it. No command, flag, environment variable, or config key moves.
Schedules
- A missing backup remote is a skip, not a failure. On a hosted intern with no backup remote, the backup schedule still fires every 30 minutes and records a run with the summary
{"skipped":"no_backup_remote"}instead of an error.GET /backup/statuskeeps returninglastSuccess: null, lastFailure: null, which the dashboard reads as no backup yet. The next fire after a remote is configured backs up normally without a restart.ori backuprun by hand with no remote still exits non-zero with the same message and hint. Aschedule_runevent withoutcome: errornow carrieserror_class.
Slack
- Slack turns act as the sender again. On an identity-layout intern, a request such as
curl https://www.googleapis.com/...inside a turn a person started gets that member’s grant injected instead of aVAULT_GOOGLE_ACTING_USER_MISSINGrefusal. Schedules and bearer-only invokes still act as nobody. A forgedx-ori-turn-credentialrefuses the write with401instead of being ignored.
A machine can pin Ori credentials to one OpenRouter organization or workspace, and an intern can act under a verified identity through the vault sidecar, the daemon, and the Slack surface. A forked session can merge back into its parent, the history store stays bounded, and the daemon asks for its bearer on every read. Slack gains a live status message, image generation through zero-data-retention endpoints, a link back from spawned threads, and cancellation when the triggering message is deleted. The CLI adds
ori alias --shim, a feature source in ori.md can name the features it loads, a skill can declare the environment it needs, and the exported telemetry gains redaction and export counters.Managed authentication
- Pin a machine to an organization or workspace.
ORI_REQUIRED_ORGANIZATION_IDpins the machine to an OpenRouter organization, andORI_REQUIRED_WORKSPACE_IDnarrows the pin to one workspace UUID. Ori reads the keys only from the user managed plist, the device managed plist, and/etc/ori/config.json. The same key in a shell, user file, workspace file, or flag is ignored. Ori combines the policy across every administrator source. A workspace pin narrows an organization pin, and two different pins block until an administrator repairs every source. Nothing changes until an administrator deploys one of the keys. - Login and the shared gate verify the key owner.
ori loginannounces the policy, sendsrequired_workspace_idto OpenRouter when a workspace is pinned, and verifies the issued key’s owner before writing anything.ori login --with-keyno longer saves a key whose verification failed under an active policy.ori code,ori dev,ori start,ori eval, andori initverify the key and fail with a closedORI_CLI_MANAGED_AUTH_*code before spending it. A personal key fails even when its creator is a member. - The native harness checks before every request. A refusal ends the run with
ORI_MANAGED_AUTH_POLICY_REFUSEDand the terminalmanaged_policy_refusedreason. Ori never retries it or switches to a fallback model. External launchers such asori claude,ori codex, andori dshcheck the credential against the policy before launch; they are not categorically blocked on managed machines. Ori does not enforce the policy inside the external process after launch. ori authreports the verdict. The command prints the policy state, its sources, the required identifiers, and the compliance verdict. It exits non-zero under a policy it cannot satisfy. The ownership line readsverified,verified, not the required owner,not verified, ornot checked. The JSON envelope gainsmanagedPolicy,credentialIdentity, andpolicyStatus, and a refusal reports its closed code such asORI_CLI_MANAGED_AUTH_ORGANIZATION_MISMATCH.- Endpoints under a policy.
ORI_OPENROUTER_BASE_URLmust be anopenrouter.aihost over HTTPS. Ownership is always checked againsthttps://openrouter.ai/api/v1/key, and the lookup never follows a redirect with the bearer token.
Configuration
- Select features from a source. A
featuresentry inori.mdcan be an object withsourceandinclude, andincludelists the exact feature directory IDs to load from that source. A bare string entry still loads everything from its source, and local paths accept the same syntax. Globs are not supported. Restart to change the selection, and runori init .to refresh remote package links. A filter does not disable features from other sources or builtins.
Acting identity
- Identity mode on the vault sidecar.
ori vault-tunnelmints short-lived turn credentials whenORI_TURN_KEY_FILEandORI_TURN_PUBLIC_KEY_FILEare set.ORI_VAULT_VERIFY_HOSTandORI_VAULT_VERIFY_PORTname the verify listener, with a default of127.0.0.1:8081.ORI_TURN_CREDENTIAL_TTLsets the credential lifetime in seconds, with a default of 21600, and must be positive. The mode also readsSLACK_SIGNING_SECRET,ORI_DAEMON_TOKEN, andORI_INVOKE_PUBLIC_KEY. The mode is off until these variables are set. - Delegated verification on the daemon and the Slack surface. Set
ORI_TURN_VERIFIER_URLandORI_TURN_PUBLIC_KEY_FILEtogether to verify callers through the sidecar.SLACK_SIGNING_SECRETthen becomes optional for the Slack surface. Setting one variable without the other answers503on every daemon write route. The daemon never falls back to local verification in that state. - Signed dashboard callers. The platform names a dashboard caller with
x-ori-invoke-signature,x-ori-invoke-timestamp, andx-ori-invoke-user. The signature isv1=followed by a base64url Ed25519 signature overori-invoke-v1:<agentId>:<timestamp>:<user>:<sha256 hex of the body>. A request that presents theORI_DAEMON_TOKENbearer is still admitted, as nobody.
Observability
ORI_LOG_LEVELgoverns the audit export. An audit event below the configured level no longer reaches the OTLP exporter.ORI_LOG_LEVEL=debugbrings the streaming deltas back on demand. At the defaultinfolevel a turn exports about 24 records instead of 104.- High-volume events export at debug.
item.started,item.updated,item.completed,tool.progress,tool.started,tool.succeeded, andtool.result.succeededleave only atdebug. They keep their own level on the hub line and on the stdout mirror thatori devreads. A record that leaves only atdebugcarries DEBUG severity, so a downstream severity filter reaches exactly the records the default level withholds. No failure event is demoted, and every event stays reachable at some level. - A cancelled run reads as cancelled.
ori.run.outcomeon anori.runspan can becancelled. Such a span ends withotel.status_code OKandstatus.interrupted trueinstead of an error. - One-shot commands export a run span.
ori code,ori eval, and any other command that boots a runtime and exits now export anori.runspan. Before, they exported none. - Incomplete runs are named.
ori.run.incompletemarks a run whose process went away while a turn was still open. It is distinct fromori.run.evicted, which still means a run leaked for a day. - A failed tool leaves a diagnostic record. A
tool.failedevent exports one ERROR record with the tool name, a classifiedcodesuch asunknown_tool,durationMs,outcome,outputChars, andtoolCallId.tool.succeededfollows the same contract. The record counts the output length and never carries the output itself, and a runtime warning getsdetailChars, neverdetail. A field the event does not carry is omitted. The record exports at the defaultinfolevel, so a feature author sees a failed script without setting anything. - Attributes export redacted. A host path or credential shape in a log attribute, span attribute, span event, status message, or metric label now exports as
<user-path>or<secret>, the way the log body already did. Identifiers, codes, model slugs, and span names are untouched. - Export health has counters.
ori_otlp_export_attempts{signal},ori_otlp_export_failures{signal,reason}, andori_otlp_records_rejected{signal}are written on every export, including by zero, so a healthy process publishes them at zero instead of leaving them absent.ori_journal_entries_missed{subscriber}counts gaps on the bounded live tail.
Sessions
- Merge a fork back into its parent.
ori code --resume <parent> --merge-from <fork>folds the fork’s committed work into the parent session and runs the prompt with it. The merge is written to.ori/session-history.sqlitebefore the model request, so the next turn in a new daemon process still has it. A fork that keeps working after a merge can be merged again and contributes only what it added since. - A merge is refused, not duplicated. Naming a session that did not branch from the resumed one, or one whose work is already merged, fails with
ORI_SESSION_HISTORY_MERGE_REFUSEDbefore any model request.--merge-fromon a launch that opens the TUI is refused too. - Schema version 3 is one way. A version 1 or version 2 store migrates on open, in one transaction, without modifying any existing node. An earlier ori build refuses a migrated store outright, so Ori copies the store beside itself as
session-history.sqlite.v2.backupbefore raising the version. To go back to an earlier build, stop ori and move that file oversession-history.sqlite. - The history store is bounded. The store keeps the 64 most recently advanced sessions plus everything those sessions reach. It never evicts a session a retained session names as its parent, and never one whose lease expired less than a further lease period ago. Before,
.ori/session-history.sqlitegrew for the life of the install. - A session with an unreadable display record still opens. Reads keep every entry that decodes and log each drop by session, checkpoint, entry index, and recorded type. The opaque payload a continuation resumes from stays strict, and every write stays strict.
- A subagent’s history survives a restart. A child thread of a session with a durable writer records its own checkpoints under a session keyed by its thread ID, so its head and checkout targets resolve after the daemon restarts.
- History errors carry the cause. A refused turn now reads
Ori could not use the session history store at .ori/session-history.sqlitefollowed by the underlying error. It tells you not to replace or delete the file on that message alone. A schema verdict names the versions this build knows, andORI_SESSION_HISTORY_STORAGE_FAILEDpoints you atori logsinstead of at your database.
Daemon
- Reads take the daemon credential. Every read route except
GET /healthnow requires the daemon bearer.ori sessions,ori logs,ori schedules,ori tui, and the chat surface already resolve it fromORI_DAEMON_TOKENor~/.ori/daemon/<port>.token. On a normal machine nothing changes. Attaching to a remote daemon needsORI_DAEMON_TOKEN=<token> ori tui --host https://<intern-host>. A refused read names the header and the environment variable instead of a filesystem path. Feature-authored routes stay open.
CLI
- Harness shims.
ori alias --shim --harness claudecreates~/.local/bin/claude-oriand prints its absolute path for a launcher such as Conductor.--dirselects another directory, and--shim --printpreviews the script without writing. Reinstallation repairs changed content or permissions.ori unalias --shimremoves managed shims. Both commands reject--dirwithout--shim, and shim operations require a POSIX host. A concurrent shim command returns a busy error while.ori-shim.lockis held. - A broken global workspace recovers on its own. When
~/.ori/globalexists but is not a usable workspace,ori codearchives it under~/.ori/backups/and rebuilds it, and the output names the backup. A second Ori process that finishes later reports that another process won and uses that workspace. There is no lock file, so a run that dies partway leaves nothing that blocks the next one.
Slack
- One live status message per turn.
status --update "<text>"rewrites the message the last--notifyposted in this turn, notifying nobody. With nothing anchored it behaves like--notify, and a failed edit posts a new message and anchors that one. The anchor is cleared when the turn ends. The prompt asks for anOn itopener with--notifyinside the first minute, one live line per action, and the final answer as the turn’s result. A turn that can just be answered still skips all of it. - Native status instead of posted acknowledgements. Every accepted request starts with temporary
On it…text, and the first heartbeat replaces it with skill prose orWorking. The runtime no longer posts timed acknowledgements, soSLACK_ON_IT_AFTER_MSis removed. The status indicator keeps the agent’s last line for the whole turn instead of reverting toWorkingafter two minutes of silence. It names an MCP server rather than its tool names, so it readsworking · Acme ×3. - Images come from zero-data-retention endpoints. The
imageskill draws on OpenRouter’s catalog of image-capable models with a zero-data-retention endpoint, newest first, refreshed hourly. It sends at most three models per request. The dashboard at/slack/dashboardgains an “Image model” select that pins one model first, andSLACK_IMAGE_MODELstill works behind the dashboard pin. A refused request reports OpenRouter’s reason instead of a bare status. - Spawned threads link back. A spawned thread’s opener carries a View original thread button whenever the origin is known.
spawn-thread newandspawn-thread forkaccept--origin-channeland--origin-tsas a pair, placed before--opener. Otherwise they read the running turn’sSLACK_CHANNEL_IDandSLACK_THREAD_TS. With no origin the opener posts without the button and stderr says why. - A deleted trigger cancels the turn. Delete your mention or DM while the intern is still working, and nothing gets posted. A steered turn with a second message keeps going, and deleting some other message in the thread changes nothing.
- The agent sees Slack profiles.
<slack_message>carries the sender’sprofilewith handle, real name, display name, and email when the app has theusers:read.emailscope, plus amentionslist for every mentioned user. - Dashboard settings save on change. Mid-turn behavior, the Cancel button, the image model, and per-thread mode save as you change them. Cancel turn, Mute, and Unmute keep their buttons.
Feature authoring
- A skill declares the environment it needs. A
requires-envlist in a skill’s frontmatter names the variables the skill needs. A run that does not carry every one of them with a non-empty value is not offered that skill. The skill leaves theskilltool, the prompt’s## Skillssection, and the slash menu. A skill that declares nothing behaves as before. The nine Slack skills declareSLACK_CHANNEL_IDandSLACK_THREAD_TS, exceptspawn-thread, which declares only the channel. The runtime logs at debug which variables a dropped skill needed.
Runtime image
- STDIO MCP servers launch. The runtime image ships node and uv, so a STDIO MCP server configured with
command: "npx","uvx", or"bunx"starts on a VM running it. No command, flag, or config key changes.
The native Ori harness now runs on a thread-first Responses loop with durable session history, forks from exact history points, and turn attachments on every chat surface. Every agent run failure carries a code that names one cause, and Datadog receives named failures, counters, and egress traces from the daemon, the Slack surface, and the vault sidecar. Login can select an OpenRouter region,
ori update applies patches instead of full downloads, and MCP servers can authorize through OAuth. Slack gains dashboard thread controls, renamed skills, and sender identity. Feature authors get declarative tool contributions and typed command arguments.Breaking changes
- Harness author API renamed.
AgentHarnessisHarness,AgentHarnessServicesisHarnessServices,AgentHarnessContributionisHarnessContribution,AgentHarnessExportisHarnessExport, andAgentInterfaceisThreadSpawner.init(host, register)becomesinit({ services, register }), andHarnessSteeringInboxbecomesHarnessSteeringReceiver. There are no old-name aliases. Refresh a generated SDK withori init .and migrate imports.AgentInput,AgentRunConfig,AgentRuntimeEvent, failures, interactions, and sessions keep their names. - Unused SDK exports removed. The main SDK drops 49 names without aliases, including persistent
ThreadHandleandThreadsCapability, theInputChannelandOutputChannelcontracts, genericAgentHook,AgentRunContextandHookType, theCompactionsprovision, and theharness-message-classificationsubpath withCreditShortfallandInvalidOutputDiagnosis. Live thread types,AskOutcome, feature hooks, normalizer APIs, and approval points remain. Chat.forkThreadisChat.forkSession.ForkThreadInputSchemaisForkSessionInputSchema. The/forkcommand, its transcript line, and the session picker now speak of sessions rather than threads.- Failure codes name one cause each. Surfaces show codes such as
ORI_MODEL_RATE_LIMITED,ORI_MODEL_UPSTREAM_FAILED,ORI_EGRESS_DESTINATION_BLOCKED, andORI_TOOL_NAME_DUPLICATEin place ofORI_MODEL_PROVIDER_ERROR,ORI_MODEL_OUTPUT_INVALID, andORI_RUNTIME_INVOKE_FAILED. Themessageis fixed Ori text. The provider body goes to the harness diagnostic log, and the HTTP status is its ownupstreamStatusfield. Every retired code in older data, including theORI_CLAUDE_*codes, decodes asORI_LEGACY_TURN_FAILEDwith its original message. /goalis no longer a built-in command. The model no longer receivescommand-goal. No migration deletes stored data.- Session resume flag.
ori,ori code, andori tuiaccept--resume <id>. A bare--resumeis a usage error. - Egress sidecar accepts full HTTP/1.1. The proxy port now decodes chunked request bodies, answers
Expect: 100-continue, forwards absolute-form request lines for the tunnel’s authority, and strips proxy-only hop headers such asProxy-Connection. Refusals answer with a framed response and a machine-readable code instead of a dropped socket. - Turn attachments.
chat.sendMessageandchat.forkSessionaccept an optionalattachmentslist. Direct API calls reject invalid content.
Native harness and sessions
- Thread-first Responses loop.
ori code --harness oriruns workspace tools, MCP tools, questions, approvals, and provider web search and fetch through the new loop when an OpenRouter key is configured. Reasoning summaries and recorded durations survive replay, and resolved interaction cards disappear. An immediately rejected harness completion with--output jsonlreports a normal failed result with exit code 1. - Durable session history.
.ori/session-history.sqlitemigrates from version 1 to version 2. An idle native session keeps its ID onori code --resume. A busy continuation branches from the last completed turn./resumeand/rewindread the same session-scoped journal, including records retained across a restart. Failed tools stay failed after a restart, and interrupted tools reload as abandoned. Unreadable history reports recovery guidance without blocking other harnesses. - Fork from an exact history point.
/fork <task>and bare/forkfork from the latest committed native history.--summaryrequests a summary fork.--from,--before, and--afterselect a point, and--beforerequires an explicit message selector. The agent can callcommand-historyto list retained points andcommand-forkto fork from one in the background.Chat.loadForkHistoryandGET /api/sessions/:id/fork-pointsexpose the same data. A rejected fork now shows the daemon’s reason in the TUI and ends an/api/invokestream with aturn.failedline. - Turn attachments. Paste a small text file or supported image in the TUI to attach it, including on
/fork. Up to eight attachments per turn, 32 KiB per text file, and 20 MB of decoded content. A PDF, SVG, HEIC, or oversized file stays a tool-readable path. Slack uploads use inline content when valid and a readable temporary reference otherwise. Under session pressure, older unpinned images become placeholders while current input stays intact. - Requests fit the egress ceiling. A turn that would exceed the ceiling replaces its oldest images with a placeholder, never the one just attached. An attachment that cannot fit fails before the request with
ORI_MODEL_REQUEST_TOO_LARGE, whose message names the encoded size and the limit. - Images a provider refused are redacted once. The harness replaces the refused images, commits a new checkpoint, and retries the request once. The
retry.scheduledevent names theredact-imagesrecovery action. Later turns start from the redacted checkpoint. - Text-only models skip images. A session on a model that cannot view images answers on that model and sees a note where the image was, instead of failing
ORI_MODEL_UNAVAILABLE. - Mixed-provider reasoning history recovers. On the mixed-history rejection,
ori codeclears historical reasoning and opaque compaction, shows a retry notice, and continues with the same model. Completed tool work does not run again. The notice discloses that an encrypted summary can lose context. - Server tool failures retry. A failed OpenRouter server tool request shows one
server_tool_failedrecovery line and retries without the server tools. - Prompt caching on Anthropic. The harness sends
cache_control, andturn.succeededusage reportscacheReadTokensandcacheCreationTokens.inputTokensis now the fresh remainder, so the three fields sum to the whole prompt. - Compaction is billed on the turn that compacts. A rollover turn’s terminal usage includes the summary run, so
/usage, session metadata, and headless cost pick it up. Each transient summary gets its own session slot, so concurrent rollovers no longer fall back to the journal projection. A steer sent during compaction reaches the turn that compaction produces and returns{"delivered":true}at once. - Withdrawn text stays out of results.
run.text(),run.output(), the Slack answer, and the eval SDK’srun.text,toolCalls, andrun.tool(name)no longer carry text or tool calls from an attempt a mid-stream retry withdrew. - No false stalls. A completed answer or tool call continues while the provider connection stays open, and a completed Responses stream settles
turn.succeededwithout waiting on the aborted reader. - MCP tool names every provider accepts. A server whose name or tool contains a dot, such as
docs.sitepublishingget.weather, is offered asdocs_site__get_weather. The server still receives its own name. ORI_OPENROUTER_BASE_URLapplies again. The native harness andori pisend requests to the host it names.turn.succeededusage listsgenerationIds, soori evalresults getservedProvidersfor native runs.- Structured output keys match the schema. A named
outputSchemaproduces a final message whose top-level keys are the schema’s own properties. The name no longer steers the model. - Restart mid-tool keeps the session. A follow-up on a session whose intern restarted mid-tool is admitted on the same session with a “Tool outcome unknown” marker, and shutdown no longer asks operators to restore a backup.
Daemon API
- Detached runs.
POST /api/invokewith"detachOnInteraction": trueends the response at the first permission or elicitation request.POST /api/interactions/respondanswers with the detached run’s identifiers, and the newPOST /api/attachstreams a receipt, the retained events, and the live tail. A run that ends abnormally sends oneattachment.endedline with a terminal reason. - Busy-session policy. An invoke may send
busySession.refusefails the second turn withORI_SESSION_RUN_IN_PROGRESSwhile the first holds its lease, andbranchor omission keeps sibling creation. A turn rejected by a running turn endsturn.failedwith that code on the invoke stream, in the journal, and in schedule history. - Caller-chosen session IDs.
POST /api/invokewith asessionIdthe runtime has never recorded creates that session. Probing for a session no longer reads the whole journal.
Failures and observability
- Every Ori failure reaches Datadog Error Tracking. Exported records carry
error.kindas theORI_*code anderror.messageas the catalog summary. New ERROR records name a degraded feature boot per skipped contribution, a failed daemon request by route family, and a failed background fiber. Retries say whether they ran out throughattemptsand aretry.failedevent, andtool.failedcarries the tool name and a code such asunknown_toolormcp_timeout. - Counters export as monotonic sums. Query every
ori_*counter with.as_count().ori_agent_runs,ori_agent_run_tokens, andori_agent_run_cost_usdgain anerror_kindtag. Theori.commandresource attribute names the launching CLI command on every exported signal. - CLI failures carry a closed code. The
cli_errortelemetry event includeserror_classandfailure_kind. Messages, hints, exit codes, and JSON output are unchanged. - Compiled stack traces name source files. Release binaries and runtime images report repo-relative files with line and column in
daemon.unhandled_rejectionrecords and the[ori-daemon]stderr line. - Vault sidecar reporting. Every completed exchange logs
vault_egress_exchange_completedand movesori_vault_egress_exchanges_attempted,ori_vault_egress_exchanges, andori_vault_egress_exchange_duration_ms. Each refusal sets anx-ori-egress-refusal: <CODE>header and writes a warn record witherror.kind. WithOTEL_EXPORTER_OTLP_ENDPOINTset, the sidecar exports onevault_egress.exchangespan per exchange and nothing else. Both proxy queues are bounded, and minted certificates are valid from one hour before minting. A headerless vault 5xx is namedORI_MODEL_OVERLOADEDorORI_MODEL_UPSTREAM_FAILED. - Chat start failures are named.
ori startin a workspace without Slack credentials logs that the surface was not started instead of a generic failure, and a start that rejects logs its error kinds. Thechat_surface_failedtelemetry event carries the error class and cause chain.
Slack
- Dashboard thread controls.
/slack/dashboardlinks threads to Slack, mutes and unmutes them, selects Steer, Queue, or the workspace default, and cancels a turn. A stale page returns a conflict instead of cancelling newer work. An optional Cancel button in Slack is off by default. - Skills renamed.
slack-statusisstatus,slack-apiislookup, andslack-ask,slack-attach,slack-chart,slack-image, andslack-questionsdrop the prefix. There are no aliases, so update pinned names, disabled entries, and script paths underSLACK_SKILLS_DIR. Compiled runtimes ship the skills, and built-in Slack exposes all nine when credentials are configured. - Sender identity. The agent receives each message’s Slack user ID and resolved name. A later interruption by another person does not change who asked, and form answers keep the submitting account.
- Acknowledgement and outcome. The first turn acknowledges with
On it: <sentence>when the model produced a complete sentence, or bareOn it. Queued and steered turns stay quiet. A turn that returns no text readsFailed: the agent returned no textinstead ofDone: no output. - Slack in Datadog. Failures export with
error.kindset to the Slack error code or the named failure class, and a bad bot token marks the surface failed at start instead of rejecting on every boot. New counters and histograms cover events, API errors, acknowledgement time, queue time, first post, and total turn time. Query@slack.event_id:<id>to find a run. Authors getChat.meterandChatTurnInput.traceAttributes. - Reliability. Queued turns keep their order across shutdown, unreadable event bodies return 400, blocker timeouts no longer read as a successful choice, and the questions route rejects an empty question ID with 400.
Authentication and updates
- OpenRouter region selection. Login discovers the regions the key allows and prompts when more than one is available, then saves
regionto the config.ORI_OPENROUTER_REGION(global,us,eu) andORI_OPENROUTER_ENDPOINToverride it, andori login --regionwins over the saved value.ori auth regionreads and sets the preference. Nine launchable harnesses receive the resolved endpoint.ompfails a non-global launch withORI_CLI_HARNESS_REGION_UNSUPPORTED. The native harness offers OpenRouter web-search and web-fetch server tools only on the global host. - Patch updates.
ori updatedownloads a sha256-keyed patch against the installed release when one exists and reports the size it saved. Complete downloads fetchori-<platform>.gzwhenSHA256SUMSlists it.ORI_RELEASE_BASE_URLpoints the updater at anotherhttpshost or a loopbackhttphost. - Alpha on every push. Every push to
mainpublishes a pre-release, soori update --alphatracks the tip. ori init .refreshes stale SDK dependencies inbun.lock, so a separatebun update oriis no longer needed.
MCP
- OAuth authorization. Add
"auth": { "type": "oauth" }to an HTTP entry inmcp.json./mcpshows consent before opening a browser, offers URL copy, cancels with Escape, and signs out withs.ori mcp test <server>and native consumers reuse and refresh the grant. Credentials live in~/.ori/auth/mcp.sqlitewith the key held throughBun.secretson macOS and Linux. Windows is unsupported. Any nonemptyORI_VAULT_*or proxy variable disables native OAuth. - Feature MCP files.
/mcpincludes enabled features’ MCP files from the daemon’s feature root. Same-name workspace and feature servers receive separate grants. Management is same-machine and daemon-cwd scoped. - Merged configuration. Global, workspace
.ori, workspace-root, and explicit--configfiles merge.ori mcp list --jsonreportsconfigPaths. MCP contract types ship fromori/mcpin generated projects.
CLI and harness launches
ori doctor <harness>replacesori harness-doctor. Binary checks resolve against the launch PATH, provider-environment warnings skip variables the harness overlay forces, and--jsonemits both reports.- Usage errors exit 2. Unknown commands, unknown flags, and invalid values exit 2 and emit one JSON envelope with a hint.
ori -V,--version, and-vprint the version instead of starting the coding agent. Runtime failures still exit 1. - Clearer help. Root help shows action summaries, and each command has its own description and
EXAMPLES. Help names model charges, approval defaults, destructive behavior, and live-runtime requirements.ori evalsummaries no longer overflow the mean duration. - Exit prints the resume command. The coding TUI prints
ori --resume "<session-id>"once on exit, omitted for a blank session after/new. - Claude launches.
ori claude --model <non-anthropic model>hits prompt caches. Seeded mode no longer disables Claude Code’s nonessential traffic, so auto-updates work while telemetry stays off. The seeded model list drops:batchvariants and models without text input and output. When the catalog seed fails, the/modelpicker showsOpus,Sonnet,Haiku, andFablewith descriptions. - Reload quiet on vanished entries.
ori devandori start --watchno longer print aNotFoundstat failure for entries that vanish or dangle. - Effect runtime update. Array-export contributions still load,
--jsonerror envelopes keep their shape,ori schedules runskeeps one row per run when a message has line breaks,ori evalreports survive a malformed generation id and a sub-cent cost, and nested tool schemas emitadditionalProperties: true.
Chat TUI
/effortconfirms once. Picking a level prints one confirmation and no next-session warning. The next prompt uses it.- Bare footer values. The footer shows session cost, session time, and reasoning effort without sigma prefixes or the effort icon.
- Logs stay off the terminal. Interactive
ori codesessions no longer get log lines painted over the input area. - Skills everywhere. Contributed skills reach the prompt and native skill tool in regular, code, and packed runs, and slash suggestions include skills from the resolved cwd.
Feature authoring
- Declarative tool contributions.
ori features new greet --kind toolscaffolds adefineToolcontribution with required input and output schemas. A handler can callspawnThread(input, parent?). Feature, surface-command, and MCP name collisions disable only the ambiguous names. MCP argument constraints are enforced before the call. - File tools reach beyond cwd.
read,write,edit,glob,grep, andread-imagework outside the working directory when the OS permits.ori code --approvals self-driveruns shell commands without the wrapper from the first attempt, and--approvals manualrestores the sandbox and approval flow. - Typed command arguments.
defineCommandtypesctx.argsfrom itsargumentsdeclaration with no type argument. An undeclared key or adefaultthat contradictstypeis a type error.
Interns on the vault architecture reach the network again. The vault sidecar addresses the vault by its own host, passes a destination’s redirect back to the agent, and stays up when the agent hangs up mid-request. A forward that still fails now answers the agent and leaves a log line.
Vault sidecar
- Outbound requests reach the vault again. The sidecar addressed the vault with the agent’s own
Hostheader, so Cloudflare rejected every forward at the edge and the intern had no outbound network. The forward now carries the vault’s authority.ori vault-tunneltakes the sameORI_VAULT_*settings as before. - A destination’s redirect reaches the agent. The sidecar refused every
3xxon the egress path, so any destination that answers a redirect failed with a dead connection, and MCP servers retried without ever connecting. A3xxthe vault relayed from the destination is written back to the agent with its status andLocationintact, and the agent’s HTTP client decides whether to follow it. A304answers a conditional request instead of failing it. The sidecar still never follows a redirect itself, and a redirect that points back at the vault’s own origin is still refused withvault_egress_redirect_refused. - The agent hanging up no longer restarts the sidecar. Writing a response to an agent that had already disconnected raised an uncaught error,
ori vault-tunnelexited, and systemd restarted it in a loop that took the intern offline. The write now fails in the ordinary error channel, and the hang-up is recorded once atinfoasvault_egress_client_gone. - A failed forward is visible. The agent gets
502 Bad Gatewaywith a stable code as the body. The sidecar writes oneERRORline that names the code, the target host, the method, the port, and a redacted reason. The vault’s response body, the agent token, and header values stay out of both.
Hosted interns back up on a schedule and restore without losing their memory. Slack gains a
!model command, a one-time progress notice on long turns, and rendered tables in status posts. The chat TUI notifies the desktop when a background turn finishes and extends caret selections across blocks. Tool calls work again on every provider, the workspace’s skills reach the model, and the eval command reports which case failed. The vault sidecar terminates TLS on the intern’s own VM and requires new ORI_VAULT_* settings.Backup and restore
- Hosted interns back up automatically. An intern that boots the plain binary with
ORI_DAEMON_TOKENset registers the builtin backup schedule, so it takes a backup every 30 minutes once a git remote exists. The Connections tab moves fromNo backup yettoHealthyafter the first fire. - The host sets the backup remote.
POST /backup/remoteon the daemon takes{ "url": "<git url>" }with the bearer credential. It acceptshttps://,ssh://, anduser@host:pathURLs, addsoriginwhen missing, and repoints it when present. The response names the remote and the change, never the URL. ori backuppushes tooriginonly. A missingoriginfails the command and lists the remotes it found. The push sets the upstream, sogit status -sbshows ahead and behind counts.- Restore matches the backup.
ori restore <sha>removes tracked files that were added after that backup. Untracked files stay.ori restore <sha> <path>still writes only that path. - Restore keeps the intern’s memory.
ori restore <ref>leaves.ori/state.sqlitealone, even when the ref predates the file.ori restore <ref> .ori/state.sqliterestores the memory explicitly. It checkpoints pending writes first and then printsRestart the intern so it reads the restored memory.When a reader keeps the checkpoint busy, the command refuses withStop the intern, then run the restore again.and touches nothing. - A rollback is a commit.
ori restorecommits what it restored asori restore <short sha>and reports it. The tree is clean afterwards, so a secondori restoreneeds no--force. A ref that already matches the tree makes no commit. ori initprepares the backup. The initial commit contains.ori/state.sqlite, the ignore file explains the!.ori/state.sqliterule, andbun.lockis no longer ignored. Interns started withori devor hosted receive a## Workspace Backupprompt section that namesori backupandori restore.
Slack
!modelchanges the intern’s model.!model <author>/<slug>switches it,!modelreports the current model, and!model resetreturns to the workspace-configured model. Slack answers the command before a turn starts, so it works while the current model is broken. It obeysSLACK_ALLOWED_USER_IDS. When the state store refuses the write, the intern says so and names the model it is still on.- A failed turn names the model. A failure the model caused now names the model, says it is unavailable, keeps the provider’s reason, and says how to change it.
- A long turn explains itself once. After
SLACK_ON_IT_AFTER_MSmilliseconds, default20000, one message lands in the thread with the model’s first sentence, or the tools it is running. Greetings, turns that have not said anything, and unfinished sentences get no notice. - Tables render in status posts. Tables, bold text, and markdown links survive in messages posted through the
slack-statusskill, matching the final answers. - Skills run wherever the package lives. The Slack surface exports
SLACK_SKILLS_DIRat startup and every skill command reads its script from it. A build without the skills fails with a sentence a person can act on.
Chat TUI
- Desktop notification for background turns. When the terminal window is not focused, a finished turn raises a desktop notification in Ghostty, iTerm2, WezTerm, and Kitty. The title is the workspace and the body is the last thing the agent said.
ORI_NO_NOTIFY=1turns it off. - Caret selection spans blocks. Shift+click extends from the caret in Ghostty and xterm without terminal configuration. A click in one message followed by a shift-click, Alt+click, or right-click in another selects and copies the whole range between them.
- Link labels are clickable. A Markdown link in an assistant reply renders its label as a hyperlink and opens the URL when clicked on the label, not only on the URL text.
Model requests
- Tool calls work again. Tool schemas no longer carry an
allOfbranch that OpenAI and Anthropic reject, so a turn that uses tools completes instead of failing withThe model provider rejected the request. - Provider rejections name the status. A failed turn reports
The model provider rejected the request (HTTP 400, InvalidRequestError).The upstream body is still never forwarded. - Retries outlast a short outage. A turn that used to fail inside half a second now retries for about eight seconds before reporting the provider’s verdict.
- Session history survives transient SQLite failures. The connection waits up to five seconds on a busy database and retries store failures. When the stored transcript still cannot be read, the turn fails with
ORI_SESSION_HISTORY_UNREADABLEinstead of answering without history. Only undecodable data marks a session unreadable, so a store failure no longer stops saving for the rest of the session. - A stable session id per run. Every OpenRouter request from a run, including task-tool children, carries
X-Session-Id. SetORI_OPENROUTER_SESSION_IDto choose it, or the runtime mintsori-<uuid>once per run. - Broken MCP servers no longer stall the others. An intern with an unreachable entry in
mcp.jsonkeeps using its other MCP servers, and the turn no longer waits 10 seconds per broken server. - The workspace’s skills reach the model. In
ori,ori code,ori dev,ori start, andori tui, the harness turn gains a## Skillssection that lists each skill’s name,/commandaliases, description, and path. The model reads a skill body with its file tools when it applies. Packed interns are unchanged.
CLI and harnesses
ori aliasintercepts native harness commands. The opt-in command writes one shell alias per launchable harness, such asclaudeandcodex, into your shell startup file.--harness <name>narrows the set,--printwrites the lines to stdout, andori unaliasremoves the block.command claudestill reaches the vendor binary.- 1M context entries in the Claude picker.
/modelinori claudelists an overridden model twice, at its capped context and with a(1M context)suffix. A model already selected asopenai/...[1m]now gets the 1M window the name promised. - Deploy images leave secrets out.
ori init --with-dockerwritesdeploy/Dockerfile.dockerignore, andori updaterefreshes it, so the image no longer bakes in.ori/,.envfiles,.git, ornode_modules. Root.dockerignorerules must move into that file. - Schedule triggers require the daemon token.
ori schedules trigger <name>sends the bearer token itself. A raw request to/api/dev/schedules/<name>without it gets401and a remediation that names the token file andORI_DAEMON_TOKEN. - Windows fixes.
ori login, including the login thatori codextriggers, completes instead of failing at the key exchange.ori initandori featuresextract GitHub archives when Git for Windows’ tar is first onPATH.ori initno longer fails while creating the workspace.oridirectory. - The vault sidecar terminates TLS on the intern’s VM.
ORI_VAULT_EGRESS_URLreplacesORI_VAULT_TUNNEL_URLand points at the vault’s egress route. The same scheme policy applies.ORI_VAULT_CA_FILEandORI_VAULT_CA_KEY_FILEare required and name where the sidecar keeps its own certificate authority. On first boot with neither file present, the sidecar mints a self-signed ECDSA P-256 CA and writes both files. On a later boot it reloads them, so a restart keeps the running agent’s trust. Exactly one of the two files present is refused. The otherORI_VAULT_*settings are unchanged. - The vault tunnel holds connections through a reconnect.
ori vault-tunnelholds a CONNECT while it reconnects instead of refusing it. It keeps retrying through401responses for a window, logsretrying, and exits with code 1 andgiving upwhen the window expires, so a supervisor can relaunch it once credentials are fixed. - Outbound requests reach the vault again. The sidecar addressed the vault with the agent’s own
Hostheader, so Cloudflare routed every forward to the agent’s target instead of the vault Worker and the intern had no outbound network at all. The forward now carries the vault’s authority. A forward that still fails answers the agent502 Bad Gatewaywith a stable code in the body and logs that code, the target host and a redacted reason, instead of dropping the connection with no response and no log line. - A destination’s redirect reaches the agent. The sidecar refused every
3xxon the egress path, so any destination that answers a redirect — an ordinary HTTP API, a GitHub endpoint, an MCP server — failed with a dead connection, and MCP servers retried on a backoff ladder without ever connecting. A3xxthe vault relayed from the destination is now written back to the agent with its status andLocationintact, and its client decides whether to follow it. The sidecar still never follows a redirect itself, so the agent token continues to reach no origin but the vault’s, and a redirect pointing back at the vault’s own origin is still refused withvault_egress_redirect_refused. - The agent hanging up no longer restarts the sidecar. Writing the response to an agent that had already disconnected raised an uncaught write error,
ori vault-tunnelexited 1, and systemd restarted it in a loop; because the agent unit waits on that listener, the whole intern went offline until the restart landed. A write to a connection that is gone now fails in the ordinary error channel, and the hang-up is recorded once atinfoasvault_egress_client_gonerather than as an incident. A502to an agent that is still connected is unchanged.
Eval
- The workspace is defined. The eval skill evaluates the directory the task prompt names, or the current directory when it names none, and treats an empty workspace as a finding rather than a reason to look elsewhere. A regrade copies the eval file instead of narrowing and restoring your committed one.
--dry-runrespects--pilot.ori eval x.eval.ts --dry-run --pilot 1previews the cases the real run would take.--pilotcredits helper modules. An eval whose case list is wrapped in a sibling helper is accepted. The refusal for an unwrapped eval no longer claims that no sibling wrapped its cases.- Reports name the failing case. The
## Failuressection ofori eval --report <path>appendson <case>to each heading when the result carries a case.agent.run({ case: "two-sum", ... })labels a run with its case. - Telemetry records the panel and the winner.
eval_runevents carrypanel_models,panel_size,conclusion,winner_model,runner_up_model,cost_pick_model, andrun_kind. ori eval docs sdkdescribes--pilotaccurately. The guard is per file, not per run or per list.
Feature authoring
- Config writes reject values JSON cannot carry.
config.writefails with an error naming the namespace and leaves the file unchanged when a value holdsundefined, a function, a symbol, aMap,NaN,Infinity, aDate, or a sparse array. Valid JSON writes the same bytes as before. - A wrongly typed contribution name is a diagnostic. A chat contribution whose
nameis not a string gets an error-level import diagnostic instead of registering under the feature id. A contribution with nonamestill defaults to the feature id, andChatContributionSchema.nameis now optional in the SDK. - Token counts and content indexes must be non-negative integers. A harness or provider that emits a fractional or negative value fails validation on entry instead of corrupting cost tracking. Sessions recorded earlier still list in
ori sessionsandori logs, with the affected figures blank.
Administrators can set every registered Ori setting through macOS managed preferences or machine policy. Schedules retry failed fires and keep their timers through a broken reload. The chat TUI adds a click caret for selecting long transcript ranges, and global flags route correctly in every form. The release also removes the
ori key-proxy command and two unused feature route context fields.Configuration and policy
- macOS managed preferences. Administrators can deploy a custom-settings payload for the
com.openrouter.oripreference domain instead of distributing/etc/ori/config.json. Ori reads/Library/Managed Preferences/<username>/com.openrouter.ori.plistbefore/Library/Managed Preferences/com.openrouter.ori.plist. The first plist that converts wins. Managed preferences sit above machine config, flags, environment variables, and user or workspace files. An invalid entry falls through to lower sources without removing valid siblings. The profile uses flat registered config keys, not anenvwrapper. - Every registered key answers to policy.
ORI_MODEL,ORI_TELEMETRY,ORI_OUTPUT,ORI_STATE_DIR, the threeORI_DISABLE_*_LOGINswitches,ORI_REQUIRE_LOGIN, and the TUI appearance keys resolve through the same chain as the update switches. Ori applies managed preferences and/etc/ori/config.jsonto the process environment before the first command runs, so the TUI, the packed launcher, and spawned harnesses inherit them. ORI_FORCE_OPENROUTER_API_KEYparses as a flag.0,false, and an empty value turn it off wherever it is set, matching the other switches.
Schedules
- Failed fires retry. A markdown schedule whose agent run fails retries up to three times, at 30, 60, and 120 seconds with jitter, before the loop waits for the next occurrence. A schedule with a custom
runhandler stays single-attempt, because the handler may already have sent something before it rejected. - Reloads keep timers armed. A feature save that boots with errors leaves the previously armed timers running, whether it produced no schedules, some of them, or could not be inspected. The runtime prints
[ori-runtime] reload armed no schedules; keeping 1 armed schedule(s) from the previous generation. A valid reload that removes a schedule still removes it. - Schedule metrics. The OTLP export gains
ori_armed_schedules, a gauge of schedules with a live timer, andori_schedule_fires{outcome=ok|error|interrupted}, a counter per fire attempt. Theinterruptedoutcome records a fire cut short by shutdown or a reload, which previously left no trace.
Chat TUI
- Click caret. A click on transcript text leaves a blinking caret at the character under the pointer. Clicking the prompt or the gutter clears it. The caret follows its text while you scroll and hides while a selection is active. Double-click, triple-click, drag, and shift-drag behave as before.
- Click, scroll, extend. After a caret click, a shift-click, Alt+click, Ctrl+click, or right-click anywhere else in the transcript selects from the caret to the click and copies the range to the clipboard. Right-click and Alt+click reach the app in terminals such as xterm that keep Shift+click and Ctrl+click for their own scrollback selection. Without a live caret, a right-click selects nothing and sets the anchor.
CLI
- Global flags route in every form.
ori --json=true versionrunsversioninstead of the coding agent.--agent=true,--json=false,--human=false,--tty=false,--no-json, and--no-humanbefore a command route the same way. ori key-proxyis gone. No deployment ran it.ori vault-tunneland vault placeholders such as__linear_api_key__are unchanged.
Feature authoring
- Route handlers lose
ctx.httpandctx.secrets. Neither was populated in a deployment. A handler callsfetchand the vault substitutes the credential.
Administrators can disable automatic updates or all updates through environment variables and machine policy files. Slack now ships the catalog implementation, with more thread actions and agent skills. Backup status comes from schedule history, and CLI errors explain why a command or model response failed.
Breaking changes
- Slack implementation replaced. The catalog’s Slack feature replaces the previous builtin. Slack remains bundled, so existing workspaces do not need to declare the catalog. A project
features/slack/still replaces the whole builtin, including its routes and skills.
Updates and installation
- Separate update controls.
ORI_DISABLE_AUTOUPDATER=1stops automatic updates and update notices while allowing manualori update.ORI_DISABLE_UPDATES=1also blocks manual updates andori update --check. Both switches override an explicit--auto-updateflag. - Machine policy takes precedence. Administrators can set the switches in the
envblock of/etc/ori/config.json, or%ProgramData%\ori\config.jsonon Windows. Machine policy wins over CLI update flags and the process environment. Workspace.ori/config.jsontakes precedence over~/.ori/config.json, and both sit below the process environment. Ori skips malformed settings individually instead of discarding valid settings from the same file. - Channel flags respect policy. A configured machine channel wins over
--alphaand--stable. An installed alpha build stays on alpha when no configuration source selects a channel. - The binary reports its version. Ori reads its embedded build version instead of an
ori.versionsidecar. The installer replaces the Ori executable already onPATH, including a versioned symlink, instead of moving the install into~/.local/bin.
Slack
- More thread actions. The bundled skills support carrying and spawning threads, attachments, charts, images, questions, status updates, and Slack API calls. Dashboard and thread routes accept only loopback callers. Slack’s signed Events API traffic still enters through
POST /slack/events. - More feature exports.
use("slack")exposespostMessage,webClient, andonButtonat runtime. The SDK still types onlypostMessage, so existing callers keep their typed API without adding a Slack client dependency.
Backup and schedules
- The state database joins backups.
ori initmakes.ori/state.sqlitetrackable by git while keeping credentials and caches ignored. Existing user rules keep their effect. Before staging files, manual and scheduled backups checkpoint the database so committed state includes recent writes when the checkpoint succeeds. A missing database or busy checkpoint does not fail the backup. - Backup status records successes and failures.
GET /backup/statusreports the latest successful and failed scheduled backups independently, within the retained run history. A success includes its time, remote name, and short commit hash. A failure includes its time and safe error detail. The route reports a read failure instead of claiming no backup ran. - Schedules can return summaries. A schedule’s
runhandler can return a flat string map that Ori stores with the run’s outcome. Existing handlers that return nothing still work. History sorts timestamps chronologically, so different time-zone offsets and timestamp precision cannot select the wrong latest backup.
CLI and harnesses
- Bad declared features no longer prevent startup.
ori startandori devwarn and skip an unavailable source declared inori.md. Ori rebuilds linked features before installing dependencies, removing stale workspace entries. Explicit--featuresarguments and the local feature root still fail when invalid. GitHub web URLs receive a hint to use the source format with@ref. - Older startup commands still boot.
ori startandori devaccept the retired--global-authflag, ignore it, and print a warning. Remove the flag from existing service commands. - Failed commands explain their exit. Ori prints one diagnostic line for a non-zero exit that the CLI has not already explained. Operators no longer see only a supervisor restart with no reason.
- Pi uses the fetched catalog.
ori piconfigures every OpenRouter catalog model, including models unknown to the installed Pi version. API selection, Anthropic cache settings, context limits, and model metadata also work in--printmode. If the catalog is unavailable or malformed, Pi falls back to its builtin models. - Vault connections retry transient failures. The tunnel retries transient HTTP 401 upgrade probes before giving up. A lasting authentication failure still terminates the connection attempt.
Chat TUI and model errors
- Image prompt cleanup. Prompt synchronization skips image-token reads after the editor is destroyed, avoiding teardown crashes.
- Invalid model output names the cause.
ORI_MODEL_OUTPUT_INVALIDexplains recognized failures, such as calling a tool that was not offered or omitting required tool-call fields. Unrecognized failures keep the generic message. The user-facing explanation never includes the raw upstream description.
Ori gains three commands:
ori muse launches Meta’s Muse Code with an OpenRouter credential, and ori backup and ori restore save and recover a workspace through plain git. The chat TUI now queues prompts by default and steers on demand. The default model moves to claude-fable-5.1, and telemetry can point at a custom ingest endpoint.CLI and harnesses
- Muse Code harness.
ori muselaunches Meta’s Muse Code through OpenRouter. Ori injects the resolved OpenRouter key asMETA_API_KEY, pins Muse’s endpoint transport to OpenRouter with bearer auth, and seeds Muse’s model catalog so any OpenRouter model id works. Muse also joins the shared harness installer. For Muse release 1.0.1-R2006.1, Ori pins a first-turn setting that keeps every turn on the configured endpoint. - Backup and restore.
ori backupstages the whole workspace, commits when the tree is dirty, and pushesHEADto the configured remote.ori restore <ref>writes a ref, or one path from it, back into the workspace, with--forceto overwrite a dirty tree. Both speak plain git and nothing else. - The default model is claude-fable-5.1. The ori harness, the init persona, and the Claude fable tier now default to claude-fable-5.1.
- Claude cache seeding matches Claude Code. Ori seeds Claude caches under the ids Claude Code normalizes to, so a seeded cache is found instead of rebuilt.
Chat TUI
- Prompts queue by default. A prompt sent during a running turn joins a queue instead of interrupting. Cmd+Shift+Enter steers immediately, and selecting a queued row steers with that row.
- The footer names the landed model. A session on
openrouter/autoshows the runtime model that actually served the turn. - Diagnostics survive shutdown. A flush race on shutdown could drop diagnostics entries. The TUI now closes the flush before it exits.
Telemetry
- Telemetry ingest is configurable.
ORI_TELEMETRY_ENDPOINTpoints the telemetry envelope at a custom ingest endpoint.
Ori can now launch Cline and Pi with an existing OpenRouter credential.
ori tui can also attach to a remote daemon over HTTPS. Vault-backed interns can call configured HTTP MCP tools, including tools that previously failed with the workspace does not grant network. Eval reports now choose a winner when the evidence supports one. The release also repairs TUI links and adds clearer runtime diagnostics.CLI and harnesses
- Managed MCP network. Vault-backed interns can call configured HTTP MCP tools. Ori grants network only when the operator provides both the vault CA and HTTPS proxy. Subprocess network access stays disabled.
- Cline wrapper.
ori clinelaunches Cline with Ori’s OpenRouter credential and selects OpenRouter as the provider. Ori seeds isolated Cline settings to skip onboarding without changing the user’s regular Cline state. - Pi wrapper.
ori pireturns as a direct launcher with model, reasoning effort,/fast, and/zdrsupport. The bundled extension forces Ori’s OpenRouter credential without writing a credential file. - Remote TUI.
ori tui --host https://<intern>.or.botconnects to a remote daemon over HTTPS. It reads the bearer token fromORI_DAEMON_TOKENand obtains workspace paths from the authenticated runtime-info endpoint.
Chat TUI
- Clickable nested-pane links.
ori devpreserves OSC 8 hyperlink sequences inside split panes. Transcript URLs remain clickable instead of rendering as dead underlined text. - Live terminal titles. Ghostty, Orca, and Herdr tabs show the workspace and current activity. The title reports when Ori is thinking, running a tool, working, or idle.
Eval
- A report states its conclusion.
ori evalselects a model, reports a close result, or explains why it declined. The Markdown report opens with the conclusion, and JSON output includes the structured decision and supporting figures.
Failures and observability
- Credit failures explain the problem. A turn that exceeds the available OpenRouter balance reports
ORI_OPENROUTER_CREDITS_EXHAUSTEDinstead ofUnknownError. The message links to the credits page and records the affordable output-token limit. - Failure details reach OTLP. Turn failures export the safe code, kind, stage, remediation, retryability, and token limit. Retry records include their attempt and delay. Production investigations can now identify the failed stage and retry state.
The pi harness is gone, and the first-party loop is the only engine. That loop now forwards workspace MCP servers as native tools, and Windows gains harness launch, install, PATH handling, and a published binary. Observability reports the outcome a run actually had and exports each log record once. Chat picks up clipboard, stdout, and Shift+Tab fixes, and a Slack run can ask real questions.
Breaking changes
- The pi harness is removed. The first-party loop is the only engine. The
ori piandharness-doctorcommands are deleted. ORI_PI_DEFAULT_MODELis renamed toORI_MODEL. The old name no longer reads.
Harnesses
- Workspace MCP servers become native tools. The ori harness reads the workspace
mcp.jsonon each invocation and registers every reachable server’s tools in the turn registry. A session can now call the Linear or Notion tools its workspace declares, instead of reporting no MCP tools. ori harness listreports install status. The command gains a launchable-harnesses section that says which harness CLIs are installed, probed with the same lookup a real launch uses. It prints the install command for a missing one, and--jsongains alaunchablearray.ori ompskips omp’s setup wizard. Ori already injects a working OpenRouter credential, so the launch now setsOMP_SKIP_SETUP=1. A first-time omp user no longer gets pushed to mint a second key.
Windows
- Harness launch works on Windows.
ori claudeand every other harness subcommand spawn the harness and forward its stdio. Installer selection, PATH persistence, Bun auto-install, and process-tree cleanup all take Windows routes, and a harness with no Windows install route says so. - The release ships
ori-windows-x64.exe. The pipeline builds and publishes a win32-x64 target with a generated nativeinstall.ps1. ori codexlaunches on Windows. The Codex provider overrides now use TOML literal strings, which the Windowscodexshim accepts.- The Windows binary ships without a smoke test for now. A Windows-only regression can ship unnoticed until the smoke gate returns.
Chat TUI
- Shift+Tab steps backward. In the
/usagepanel, Shift+Tab moves the tab selection backward, and in the composer it no longer opens the session picker. - Stray stdout cannot corrupt the frame. The TUI captures stdout during the session and routes stray writes into the diagnostics log. The exit diagnostics line prints only when
ORI_TUI_DIAGNOSTICS_SUMMARY=1is set. - Clipboard copy works inside tmux. The copy path reads the OSC 52 result and falls back to pbcopy, wl-copy, or xclip by platform, and never in SSH sessions. A copy that reaches no clipboard shows a red failure toast instead of silence.
- Expanded thinking traces render as plain prose. The Markdown renderer treated bracketed prose such as
[REDACTED]as links and underlined the leftovers. The expanded body now renders the provider’s text verbatim, and real URLs stay clickable.
Slack
ask_userrenders as interactive Slack blocks. A Slack run gets the realask_usertool. A question renders as clickable option buttons, or an answer modal for free text, and the answer resumes the run.
Sessions and interns
- A fresh session’s first turn is readable. Session-filtered reads recover the run and turn start events that land before the provider announces a session id. Rewind and checkpoints can target the first message, including in journals already on disk.
- Packed interns keep infrastructure details private. A packed intern answers capability questions in user terms. The model stops narrating proxies, environment variables, and raw status codes to end users.
Observability
- The run metric reports the real outcome. Every ori run used to land in one
cancelledbucket withmodel=unknown. The run metric now records the terminal result, andmodelfalls back to the resolved slug. A run’s trace also links to what the run spent at the gateway. - Traces stay bounded. Long-lived monitors no longer collect every journal entry under one immortal trace id.
- Each diagnostic exports once. OTLP logs no longer arrive twice.
- An event says which deployment it came from. The telemetry envelope gains optional
deployment_idanddeployment_environment, read fromOTEL_RESOURCE_ATTRIBUTES. A hosted intern run is now distinguishable from a laptop run on the adoption dashboards. - An eval run carries one session id.
ori evalresolves and locks an OpenRouter session id once and stamps it on every model, embeddings, and rerank call. Gateway records can list one run’s traffic with its cost.
Runtime image
- The runtime image is multi-arch.
ori-runtimepublishes an amd64 and arm64 manifest list, so an Apple Silicon machine pulls a native image instead of crashing under emulation.
Subagents take the largest share of the release. A parent can pick a child’s model, watch its output while the child works, and demand a schema-shaped answer. Observability is the other half. Run and turn lifecycle, diagnostics, state store activity, Slack, and daemon process health now leave the machine over OTLP. Elsewhere the goal loop spends fewer turns on the same goal, chat gains a system theme, a rewind gesture, and mid-turn steering, a feature can claim runtime behavior, and the turn step ceiling is gone.
Breaking changes
harness-customreplaces theharness-unknownfallback. Theagent_runtelemetry event always names the harness that ran. A run on a harness someone wrote reportsharness-custom. A dashboard that filters onharness-unknownmust move toharness-custom.
Subagents
- A child takes its own model, effort, name, and directory.
taskandspawn_agentaccept optionalmodel,effort,display_name, andcwd. An omitted parameter inherits from the parent. Acwdoutside the parent’s workspace root fails the spawn before any child starts. - Partial output is readable while a child runs. A new
read_agenttool returns a bounded tail of the child’s output plus a cursor. Passing the cursor back returns only newer output, and a stale cursor returns an explicit gap.stop_agentfreezes what the child produced, and a finished answer survives history pruning. - Live subagent rows in chat. A blocking
taskshows the child’s latest text on its own tool row. A backgroundspawn_agentgets a row that fills in while the child works and ends with the answer. - A declared result schema.
taskandspawn_agentacceptresult_schemaas a plain JSON Schema object. Ori validates the terminal answer only, retries once in the same child session, then fails the call. The retry feedback names issue paths and carries no values. - Late child spend reaches the session. A background child that settles after your last turn records its spend against the session. Session totals used to under-count that spend.
Goals
- A goal spends its turns on work. The seed turn can claim a goal that is already finished. The audit turn reads the session’s recorded tool calls instead of running the work again. A continued turn stops restating earlier turns.
- A claim ends its turn mechanically.
update_goalsettles the turn when it records a claim or an ending. No extra inference narrates a claim that is already saved. A refused call changes nothing and the turn keeps running. - The claim reads apart from the audit. The first
completerenders asClaimed complete, and the audit confirmation renders asConfirmed complete. Both used to read the same.
Chat TUI
- A system theme that follows the desktop. On an Omarchy desktop,
ori codereads the active desktop palette and paints itself with no configuration. It picks up a theme switch on the next launch. Every other desktop behaves as before. - Double-escape rewind. Press Escape twice while idle to pick an earlier message, then pick a restore mode. Conversation restore forks a new session from that point, so the original session stays intact. Code restore is listed and stays unavailable until workspace checkpoints exist.
- A native
ask_usertool. An agent on the default harness can ask the person a question, with optional choices, and read the answer that turn. The tool registers on attended turns only, so an unattended run cannot hang on a question. Children never see it. - Readable tool calls everywhere. The chat TUI, Slack, and the headless prose output render tool calls the same way.
update_goalno longer shows raw JSON, and an unknown tool gets a readable fallback. - Typing mid-turn steers the agent. Press Enter while the agent works and the message queues above the composer. Running tool calls continue, and the agent folds the text in at the next step boundary. A message the turn never picked up submits once as the next turn. Escape still stops the turn.
- Enter runs the highlighted slash command. The composer used to complete the command’s name instead of running it.
Turn behavior
- No step ceiling and no loop detector. A native harness turn now runs until the model finishes, you interrupt it, or the provider fails. The 64 step default killed real build turns.
/effortworks on the default harness. The ori agent loop forwards the picked reasoning effort to OpenRouter. Every level used to be refused on every model. A subagent inherits the session’s level.- Anthropic models cache on the default harness. The agent loop marks the stable prompt prefix with
cache_controlbreakpoints. The second turn of a session reads the cached prefix, and the prompt cost drops.
CLI
- Every command checks for an update first.
ori claude,ori codex, and the other interactive commands run the update gate that onlyori codehad. The three choices are unchanged.update,version,help,dev, andstartstay exempt. The persisted preference and theORI_NO_UPDATE_CHECKopt-out are shared across commands. - A feature written after boot loads. A deployed intern can arm a schedule it just authored. The deploy templates run with
--watch, and a newly added feature imports its contributions.ori schedules listcounts the schedule files on disk and reports them apart from what the daemon returns, in text and in JSON. ori omplaunches the omp harness. The command forwards the picked model and reasoning effort and injects the OpenRouter credential. Whenompis missing, ori offers the official installer, like the other harness commands.
Feature authoring
- A feature can claim runtime behavior. A new provisions contribution lets a feature supply the approval policy, the approval asker, compaction, and unattended approvals. The shipped implementation stays the fallback when nothing claims a point or a provider fails.
- A sandbox-bounded approver answers unattended asks. An unattended ask used to be denied outright when no feature claimed the point. Reads and writes inside the workspace sandbox now run, execution runs only while process isolation is active, and everything else denies. A feature-provided approver still replaces the fallback, and its deny stays terminal.
Approvals
- Escalation guidance names a refused read. The bash tool said the sandbox refuses a write or the network. A model therefore never retried after a refused read such as
git statuson.git/config. Both strings name a read now. The self-drive copy states what the mode does, which is approval without prompting and a start inside the sandbox. A refused command may retry outside it.
Observability
- Runloop lifecycle over OTLP. Set
OTEL_EXPORTER_OTLP_ENDPOINTand run, turn, and session lifecycle export as OTLP log records. Each record carries how late it was reported. A silent run raises one stall warning, and heartbeat, uptime, and active-run metrics report daemon liveness. - Diagnostics on the same pipe. State store writes, failures, slow reads, and database size export over OTLP. So do harness subprocess spawn, exit, timeout, and kill records, and the Slack boundary’s event receipts, API retries, and failures. Message text, prompts, and arguments stay out of every record.
- Process health gauges. The daemon samples
ori_process_rss_bytes,ori_jsc_heap_size_bytes,ori_jsc_heap_capacity_bytes, andori_event_loop_delay_msevery 30 seconds. Crash markers, child process gauges, audit export, and feature exec diagnostics fill the remaining gaps. A finished turn no longer sits in the active-run gauge and no longer fires a false stall report. - Every signal carries the service version. A released binary stamps
service.versionon every exported trace, metric, and log record. - A failed run names its failure class. The
agent_runandschedule_runtelemetry events carryerror_classwith the classified failure code. An exhausted balance no longer reports the same as a context overflow. - Daemon logs reach
ori logs. Daemon logs route through the log hub, soori logsshows them at last. - An undecodable response says what failed. A response Ori cannot read now maps to
ORI_MODEL_OUTPUT_INVALIDwithkind=protocolinstead of naming the provider. The reason that names the failed check reachesori logs.
Eval
- A run reports what it did. An
ori evalrun names the model it ran and prints a cost you can read. It also names the provider that served each chat row. The scratch workspace it leaves behind runs straight away.
Installation
- The Homebrew tap hears about a stable release. A published stable release notifies the tap, so the formula stops going stale between scheduled runs.
Goals stop leaking across projects and sessions. The usage panel says where the money went, and a mistyped model slug heals against the OpenRouter catalog instead of failing the run.
Goals
- Each project keeps its own standing goal. A new
ori codesession in any directory used to read the one global goal row. The daemon now keys the goal row by the launch directory, so every project holds its own goal. - A done goal renders only where it ran. Once a goal finished, the footer kept showing its green
completechip in every later session. The chip now appears only in the session that watched the goal run, and/goalstill reports the true status on request.
Chat TUI
- Cost attributed by source. The usage panel Session tab gains a per-tier cost column next to the token counts. The column apportions the billed cost with catalog list prices as weights, so the parts add up to the Cost row, and the Cost row moves to four decimals to match.
- A named cold-cache row. The Overview tab reports what a cold cache cost in a
Cache coldrow. Expiry, a provider switch, and a changed prompt prefix all produce the same token counts, so the row names the state and not a guessed cause.
CLI
--modelslugs heal against the catalog. A close but wrong--modelslug on a harness run resolves to the matching OpenRouter catalog model, with a warning that names the substitution.ori codeheals its model once and hands the healed value to the TUI, so the launch path fetches the catalog a single time.
This release is chat TUI polish. The model picker resolves dynamic variants, prints readable prices, hides batch rows, and gains standard keyboard navigation. The composer and the slash menu line up with the transcript, and a theme file can now replace the built-in palettes.
Chat TUI
- Variant models resolve their metadata. A
:nitroor:floormodel used to behave like an unknown model: wrong effort options, a spurious unvalidated-catalog notice, and no effort adoption. Model lookups now fall back to the variant’s base id, so/effortand model pick treat the variant like its base model. - Readable picker prices. Prices round to at most two decimals, and a paid rate that would round to zero clamps to 0.01 so it never reads as free. An unpriced row shows a dash instead of putting its context length under the price header.
- No batch variants in the picker.
:batchendpoints are not usable interactively, so the picker hides their rows. An explicit/modelwith a typed:batchid stays valid. - More picker navigation keys. Pickers accept
ctrl+nandctrl+pfor next and previous,tabandshift+tabto cycle, andpgupandpgdnto move a page at a time. - The composer column matches the transcript. Typed text used to start two columns right of a submitted user line, so the text jumped left on submit. The caret and the text now share the transcript columns, and the slash menu and image-path rows shift to match.
- A cleaner slash menu. Command descriptions drop the parenthetical UI narration and use the real terminal width instead of a fixed 56 columns. The tab hint stays on one line at narrow widths.
Themes
- Custom palettes from a theme file. Point
tui.themeFilein the feature config, or theORI_TUI_THEME_FILEenvironment variable, at a palette file to restyle the whole TUI. Saving other appearance settings preserves the configured file.
Approvals settle on two modes, and the classifier that used to judge asks for you is gone. Eval takes the largest share of the release. It can now bake off embedding and rerank models, grade against examples you supply, and run without a project’s persona leaking into it. The rest is repair work in the Slack egress path, the dev container, and the container image publish.
Breaking changes
self-drivereplacesbypass, andautois gone. The default mode approves every command ask for the session and never prompts.--approvals bypassand--approvals autostop parsing, so a script that passes either must drop the flag or passself-drive.
Approvals
- Two named modes. Run
ori code --approvals self-drivefor standing consent, orori code --approvals manualto answer every ask yourself. Bare/approvalsin chat opens a two item picker, and chat prints the active mode when a run starts. - Nothing judges an ask for you. The classifier that answered attended asks is gone. A reviewing mode comes back once it is evaluated, or as a declarative policy that a feature registers.
Eval
- Embedding bake-offs.
ori evalcompares embedding models end to end. Discovery reads the embeddings catalog directly, so the chat model picker keeps its own filter. The report now renders nine decimals, because a single embedding call used to print as$0.000000. - Rerank bake-offs.
setupRerank()grades rerank models on the same recorder, with NDCG@k and reciprocal rank matchers. - Reference examples for the judge.
setupJudgetakes areferencesfield that reaches the judge as its own prompt section. Your criteria still decide how much those examples count. Setup accepts at most 10 entries and 20,000 characters. Over either limit it refuses the corpus and names what it got, rather than grading against a truncated version of it. - Hermetic runs.
ori eval --hermeticbuilds a temporary root and links only the features that would have loaded. A project’sori.md, agent guides, and skill roots stay out of the run, and Ori materializes its skills into the temporary root. A run without the flag behaves as before. - No fake effort comparisons. Eval carries the harness it resolved into the run. When that harness cannot express the effort levels you asked to compare, it says so by name. It used to run one level several times and present the results as a comparison.
- No token clamp. The eval path never clamps
max_tokens, so an earlier cap cannot cut a long answer short. - An affordable pinned judge. The default judge moves to
openai/gpt-5.6-sol.
Bug fixes
- Slack through the egress proxy. Slack Web API calls follow the configured egress proxy, so a deployment that allows only proxied traffic can post again.
- Dev container on Bun 1.4. The dev image moves to Bun 1.4, and init no longer crashes inside it.
- Container images publish again. No container shipped for 0.8.0 or 0.9.0. The runtime image publishes with this release.
This release renames the never-prompt approval posture to
bypass and makes it the default, so auto keeps its classifier middle ground. Login now starts with an explicit chooser, remembers what you picked, and lets an enterprise switch off any sign-in method. Schedules arm while the daemon runs, so a cron edit takes effect on the next reload instead of the next boot.Breaking changes
bypassis the default approval mode. A default session answers every ask yes, including the escalation ask a refused shell call raises. The chat TUI announces the active mode when it starts.autokeeps the classifier. A classifier answers routine attended asks. A doubtful ask or an escalation still reaches you, and the declared ask-on-execute policy stays in force.- The credential-gate scope modes are gone. The
--global-authand--no-global-authflags are removed. Login mode now decides where a credential comes from.
Approvals
- Three named modes. Run
ori code --approvals bypass|auto|manualorori tui --approvals ....manualprompts on every ask. - Mode picker in chat. Bare
/approvalsopens a picker for the three modes./approvals manualstill sets the mode directly.
Login
- Login chooser. First run asks how Ori connects to OpenRouter: OAuth, a pasted API key, or the
OPENROUTER_API_KEYin your shell. Ori no longer adopts a key it finds in your environment without asking. - Persisted preference. The choice lands in
~/.ori/config.jsonasloginMode, so the chooser stays out of the way. Runori loginto see the current mode and pick another one. - Enterprise disable flags. Set
ORI_DISABLE_OAUTH_LOGIN,ORI_DISABLE_API_KEY_LOGIN, orORI_DISABLE_ENV_KEY_LOGINto remove a sign-in method. The chooser, the hints, and credential resolution all respect the flag. - Key validation. Ori validates a pasted key before it saves it and reports a failure in plain language.
Schedules
- Live arming. A feature reload re-arms the schedule timers under
ori dev,ori start --watch, and the split session. A cron edit used to write a correct file that no timer ever read. - Honest state in
ori schedules. Each schedule reports anarmedfield that reflects a real timer. An unarmed schedule reports no next fire time, so a disabled schedule, an invalid cron, and a schedule with no upcoming fire all read as not armed. - Guidance for agents. The schedule skill states that a schedule arms after the turn that writes it ends, so an agent stops polling
ori schedulesfor a reload that cannot land yet.
This release makes Ori’s own agent loop the default harness and unbundles Pi. The loop ships built-in tools, steering, background subagents, session resume, and an approval posture that answers routine command asks for you. Threads let a feature start child work, watch its events, intercept its operations, and ask a model for typed output. The CLI adds MCP and harness inspection commands, starts faster, and asks you to confirm a key it found on disk. The chat TUI reports usage and spend, and it recovers when your OpenRouter credit runs out.
Breaking changes
- First-party loop by default.
ori codeandori tuirun on Ori’s own agent loop instead of Pi. Launch another harness explicitly when you need it. - Pi is no longer bundled. The CLI does not ship or install Pi. This release still carries the Pi fixes below for installs that keep it.
- Key confirmation. Ori asks you to confirm before its first billable call when it uses an OpenRouter key that it discovered in a project directory. Non-interactive callers that relied on the previous silent behavior must approve the key in advance.
Agent loop
- Native turn loop. The loop runs turns in process on the OpenRouter API, with steering, an ordered message inbox, background subagents, and goals.
- Built-in tools. The loop ships its own bash, read, write, edit, glob, and grep tools.
- Sessions. The loop persists session history and resumes a previous session.
- Compaction. Long sessions stay inside the context window.
Approvals
- Auto approvals by default. An attended session answers routine command asks with a classifier verdict on the session model. Doubtful, destructive, or escalated calls still reach you. Pass
--approvals manualto answer every ask yourself. - Approval and sandbox axes. The native harness separates who approves a call from what the sandbox allows it to do. A remembered answer outranks the classifier in both directions.
Feature authoring
- Thread primitive. A feature creates threads with parent links and a spend budget. Cancelling a parent cancels its children.
- Live event view.
watchgives each thread a bounded live view of its own events. - Operation chain.
interceptgives each thread a typed chain that can observe and change its operations. - Structured output.
promptForasks a model for a value that matches a schema, so a feature thread receives typed output. - Extension points. Extension points now use typed tokens and declare a conflict policy, so two contributions to one point resolve predictably.
- Compaction hook. A compaction extension point lets a feature supply its own strategy.
CLI
- MCP inspection.
ori mcp listshows the configured MCP servers, andori mcp testchecks that one of them answers. - Harness inspection.
ori harness listshows the harnesses the CLI can launch. - Faster cold start. The CLI starts about 300ms sooner.
- Update visibility.
ori updateprints the current version and the target version, and it callsbrew upgradefor a Homebrew install. - Key ownership. Login says whose key Ori is about to spend.
- Model picking.
ori claudeseeds capabilities per model, the picker drops catalog models that cannot take text input, and first-party Claude rows read cleanly. - Workspace model.
ori codeinherits the project workspace model, andori harness testpicks its own model and passes on a fresh workspace. - Scaffolding and validation. Feature scaffolding produces working hooks and API features, and the feature validate summary agrees with the flags you passed.
- Harness binaries. The CLI resolves harness binaries against the live PATH and keeps the installer’s bin directory on it.
- Pack and schedule. Pack keeps stripping pointer keys after metadata empties, and the scheduler rejects a cron expression that it cannot arm.
Chat TUI
- Usage panel.
/usageopens a panel that reports usage and spend for the session. - Credit recovery. The TUI opens the model picker when your OpenRouter credit runs out, so you can continue on another model.
- Readable output. Clickable links stay underlined at rest, tool arguments share one clipping budget with their disclosure, and console diagnostics go to a file instead of the frame.
Pi
- Prompt caching. Pi routes Anthropic models through the completions path, so prompt caching works again.
- Token caps. Pi marks and releases the max-tokens cap that a 402 response forced, and it clears caps that earlier versions left on disk.
- Setup checks. Ori enforces a minimum Pi version and adds a Pi check to the harness doctor.
- Session and extensions. An inherited OpenRouter session id wins in the Pi runtime, the MCP extension loads, and a 402 failure keeps the full keys-page URL.
- Setup warnings. Pi setup warnings no longer print into the frame.
Eval
- Failed runs. Eval stops calling a model after three consecutive failed runs, and it never sends an empty candidate to the judge.
- Judge provenance. Eval pins the judge slug it resolved and records it with the run.
- Report scope. The report states the tallied scope and reconciles it against the durable record.
- Context inventory. Eval prints a context inventory before the run spends anything.
Runtime
- Clean stdout. The daemon keeps its stream diagnostics off stdout.
- Prompt order. The system prompt places the persona first and the dashboard instructions last.
- Pipe safety. An unread prompt pipe no longer fails a harness run with EPIPE.
Skills and Slack
- Log inspection. The code feature ships an inspect-logs skill for cross-session diagnostics.
- Self-development skill. The feature-development builtin skill is now called selfdev.
- Slack voice. A workspace persona outranks the surface voice rules.
- Slack tables. Markdown tables convert to a form Slack accepts instead of failing the post.
This release adds the Prime Agent, DeepSeek Harness, and Grok Build harnesses to the CLI.
ori pi now runs through the OpenRouter Responses API, and one flag pins reasoning effort for every harness. The chat TUI lists your OpenRouter presets in the model picker and reports per-turn speed. The runloop also fixes several skill materialization faults, and macOS release binaries are signed.Harnesses
- Prime Agent.
ori prime-agentlaunches Prime Agent through OpenRouter, andori primeis a shorter alias for it. - DeepSeek Harness.
ori dsh setupprepares DeepSeek Harness, and the setup output now points atori dshinstead of a baredshcommand. - Grok Build.
ori groklaunches Grok Build through OpenRouter. - Pi transport.
ori pisends its turns through the OpenRouter Responses API. The Pi model picker hides builtin providers that OpenRouter does not serve. - Reasoning effort. One
--reasoning-effortflag pins the thinking level for every harness the CLI launches. - Documentation. The CLI docs describe the DeepSeek Harness and Prime Agent commands.
CLI
- Link control.
ori unlinkremoves a previous link, andori link --no-pathlinks without touching your PATH. - Install prompt. The harness install offer defaults to yes, so Enter installs the harness.
- Pi repair. The CLI reinstalls a Pi runtime whose install never finished.
- Traffic attribution. Each harness launch attributes its OpenRouter traffic to that launch.
Chat TUI
- Preset picker. The
/modelpicker lists your OpenRouter presets, and it aligns model prices into readable columns. - Turn speed. The usage row reports tokens per second for each turn.
- Quieter prompts. The
ask_userprompt renders with less noise, and a folded group holds its command row at narrow widths.
Runtime
- Skill materialization. The runloop stops generating duplicate-name skill materializations and revalidates a reused skill snapshot.
- Progress reporting. Materialization progress stays isolated per command.
- Command output. Audit output no longer blocks a command that writes to stdout.
Docker and secret vault
- Placeholder warning. The vault stub warns when a resolvable placeholder rides in the request body.
- Vault root rotation. Docker restarts the containers that trust a re-minted vault root.
Slack
- Bounded queues. A Slack thread bounds its own turn queue, so one busy thread cannot queue unbounded work.
Installation
- macOS release binaries. The release pipeline signs the darwin CLI binaries before it publishes them.
This release adds a local vault stub that terminates TLS and injects secrets for approved destinations. The CLI can browse curated release notes, launch Pi, and report selected entries before it exits. The chat TUI shows turn, tool, and session durations.
Breaking changes
- Eval JSON tags.
data.pilotinori eval --jsonnow includes a_tagwith the value"report"or"notice". Consumers that read this field must handle the tagged shape.
Docker and secret vault
- TLS and secrets. The dev vault stub now terminates TLS, mints destination certificates, and injects
__{name}__values from thevault__{agentId}__{name}namespace before it calls an approved destination. - Destination controls. The stub refuses blocked destinations during a stream and serves its tunnel over secure websockets. The destination list and secret values now live in a client-editable file.
- Connection safety. The stub keeps the vault root key away from the agent, preserves binary request and response bodies, and forwards safe destination headers. It also removes injected credentials from response data.
- Workspace routing. The vault tunnel upgrade sends the
x-workspace-idheader so the service can route the connection to its workspace. - Service health. The vault tunnel and upstream now expose healthchecks, so Docker can wait for both services before it starts dependent work.
CLI
- Changelog browsing.
ori changeloglists curated releases and supports version selection,--all, and--list. A selected entry prints and exits without opening an interactive prompt. - Pi command.
ori pistarts Pi with Ori’s OpenRouter environment and passes remaining arguments to Pi. - Claude environment. The Claude command clears conflicting provider settings and applies Ori’s OpenRouter environment before launch.
- Eval outcomes. Eval derives each run from its terminal event, so final reports use the completed outcome.
Runtime
- Snapshot publishing. The publish lock no longer blocks concurrent snapshot work.
Chat TUI
- Duration display. The chat TUI now shows turn, tool, and session durations. It preserves timing metadata when usage is absent and keeps duration columns readable at narrow widths.
This release removes harness switching and gives the main ori code one harness engine. It routes agent egress through an authenticated secret vault tunnel, so a feature can reach an upstream service without ever holding the credential. Eval can price a run before it starts, sweep reasoning effort, and take a chosen judge, and it no longer scores a crashed run as a wrong answer. Slack streaming reports liveness, reasoning, and staleness.
Breaking changes
- One harness engine. ori code removes harness switching.
Secret egress
- Vault tunnel. Agent egress travels to the secret vault over a websocket, and the tunnel’s CONNECT listener authenticates its caller.
- Redaction. Redaction of substituted headers covers the whole outbound pipeline, and the key proxy reports its own defects as its own failure instead of an upstream 502.
- Stream close. A graceful stream close delivers the bytes still buffered behind it.
Eval
- Cost control. The
--pilotflag measures what a run costs before you commit to it. - Judging. A comparison asks which judge grades it, and eval warns when the judge shares a model family with a candidate.
- Sweeps. A comparison sweeps reasoning effort as well as models, and the eval docs prescribe a concurrent candidate sweep.
- Honest results. A run that died is reported as a failure instead of a model that answered wrongly. Assertion failures stay readable, a dry run names the real top level throw instead of blaming imports, and markdown table cells escape backslashes before pipes.
Slack and chat
- Streaming signals. A streaming turn sends a liveness heartbeat, reports reasoning status, and marks itself stale when it stops progressing.
- Dead runtime. The chat surface ends when its runtime dies mid turn instead of waiting on a turn that cannot finish.
CLI
- Troubleshooting.
ori harness-doctor codexdiagnoses Codex install conflicts.
This release fixes what long sessions got wrong: reported spend, session resume, and journal volume. Feature authors get a secret handle they can pass but never read, and the eval and test SDKs are now authored TypeScript. The CLI cleans up its help surface and its launch update prompt.
Runtime and sessions
- Spend across processes. A session keeps the spend recorded by earlier processes when a terminal state lands, so the reported cost covers the whole session.
- Resumed sessions. The inbound record limit rises past the size that broke session resume.
- Journal volume. Per token runtime events no longer flood the journal.
- Stall detection. The runtime reports a stall that happens before the harness starts.
Secrets and SDKs
- Secret handle. A feature can receive a secret handle, pass it to a capability, and never read the value behind it.
- Eval and test SDKs. The
ori/evalandori/testsurfaces ship as authored TypeScript instead of generated shims.
CLI
- Help.
helpis a registered command, and the root listing shows a narrower set of commands. - Launch updates. The launch update prompt offers “Always auto-update on launch” as its first choice.
Documentation
- Release notes. Curated release notes live in
docs/changelogs, one file per version. The release mirror states a security contact and reporting path.
This release makes headless and non-Anthropic use practical. Headless runs return agent questions to their caller, login works without a browser, and Claude uses credentials stored separately from the user’s. An active log file rolls without interrupting followers. Evals report the provider behind each model and the reason a candidate was rejected, and the CLI release ships a filtered source tarball.
Breaking changes
- Headless questions go back to the caller. A headless run returns an agent question to whoever invoked it instead of answering it itself.
Credentials and harnesses
- Headless login. Login no longer requires a browser handoff, and
--with-keyaccepts a key directly. - Claude credential isolation. ori neutralizes conflicting
settings.jsonenvironment values and keeps Claude’s credential store separate from the user’s. - Non-Anthropic models.
ori {harness}handles non-Anthropic models and stored credentials. The default no longer setsENABLE_TOOL_SEARCH, which lets those models work. The Claude path prefersANTHROPIC_API_KEYover Bearer auth.
Evals
The eval report names the provider that served each model. A rejected candidate now includes the condition that rejected it. Discovery failures are surfaced instead of reported as an empty eval list. Tool support is named when no endpoint serving the model provides it.Logs
The active run file rolls while a follower keeps reading across the roll. Log lookup finds the event log a run just wrote.Sessions
Session ownership persists outside a feature workspace. A session’s metadata sidecar continues across processes. The session clock starts at process start.CLI and install
- Diagnostics. A boot warning about the global workspace now points at
ori workspace reset. The same hint appears on the other boot warning path. - Correctness. Unknown subcommands error instead of falling through to
ori code. The summary line reports prompt tokens rather than the uncached remainder. - Installer. The installer refuses to run under a non-bash shell.
Release and Slack
The CLI release includes a filtered source tarball. A Slack attachment reaches the agent as a URL it can fetch.This release moves harness authors onto a provider-neutral agent connection seam and lets Claude compact context in place. It also adds native Codex support. The chat TUI and eval workspace handling receive focused fixes, and smaller fixes tighten harness defaults and Pi turn errors.
Breaking changes
- Agent connection seam. Harness authors register live provider connections through a provider-neutral runtime seam. The runtime owns session identity, resume, interruption, cancellation, and interactions without knowing the provider protocol.
- Harness initialization and compaction. Harnesses now initialize from author definitions, and Claude can compact conversation context in place.
Codex
- Native Codex support. Codex runs as a native harness, and its events project into the runtime’s event stream.
Bug fixes
- Chat TUI. Glider frames are deduplicated for a seamless loop.
- Eval workspaces.
create-evalcollects the workspace once and reuses the record. - Claude and Pi harnesses. Claude enables
ENABLE_TOOL_SEARCHonly for Anthropic models. Pi now fails the turn when the assistant emits a trailing error.
This release turns Headless
ori code into a real headless tool. A prompt now runs the agent to completion and prints a result instead of seeding an interactive chat, and callers that need to parse the output get structured JSONL. Sessions also outlive the process that started them, so a Claude or Pi run can be resumed after the CLI that launched it is gone. Alongside that, the chat TUI reworks how live activity and tool calls are rendered, ori eval learns to compare providers and validate an eval without spending model calls, and the source ships under Apache-2.0.Breaking changes
- A prompt runs headless. Passing a prompt to
ori coderuns it to completion and prints the answer instead of seeding the chat. - Code interaction forwarding is gone. The forwarding path that let a headless run hand its questions to an outer caller is removed, and
create-evalno longer answers its own questions.
Headless ori code
ori code runs headless when you pass -p and there is no TTY, and it can emit structured events with --output jsonl alongside --print. A finished run prints a summary of what it did.Chat TUI
Live activity is now bracketed as a group with the command in hand previewed while it runs, consecutive thinking and tool rows collapse onto one line, and the detail cell on a row is chosen per tool category. An expanded tool call is framed as a card holding both the command and its result. The working indicator glides.The workspace row was redesigned and its placement is configurable. The model picker prices each row and resolves aliases inline, and reasoning effort is configurable from that same picker. Selection got better: shift-click extends a transcript selection, double-click selects a word and triple-click a paragraph, and transcript links highlight individually on hover.Evals
ori eval can compare the providers serving a model and recommend one to pin, backed by recording which provider served each run. A --dry-run validates an eval without spending model calls, and human output shows scores and the models they resolved to. ori eval scratch runs self-contained temporary evals, ori eval skill makes the eval-authoring skill reachable, and the eval reference moved behind ori eval docs. Eval files may no longer use absolute imports.Sessions and harnesses
Session ownership is now persisted durably, which lets a Claude session be rebuilt from its ownership record and a Pi session be resumed after the owning process is gone. The Claude harness reports token usage and cost. Pi bash calls can carry a human summary.Runtime, telemetry, and licensing
The ori source is licensed under Apache-2.0, as is the public ori-releases mirror. Telemetry emitssession_start and session_end for sessions that run agents. OpenRouter traffic is attributed to the ori session that produced it, and terminal usage is accounted per turn. /fork runs as a concurrent background session. Skills gained a metadata.command-alias, which gives create-eval a /eval alias. The key-proxy scrubs substituted secrets out of the response leg, and the OAuth callback pages picked up a terminal look in the OpenRouter theme.Credentials
ori auth reports the resolved credential. When a project dotenv key and an ori login credential both exist, ori asks once per workspace which to use.Bug fixes
Chat TUI. Only the status line animates while a turn runs, the glider indicator loops as a seamless stream, and elapsed time in it reads in human units. Abandoned tool calls are marked instead of sitting pending forever, activity groups start folded, clipped tool args expand with their tool call, expanded args survive on labeled rows, and edited file rows color their added and removed counts. An answered elicitation shows in the transcript, a free-text Ask User prompt receives typed keys, and a projected working directory is quoted with CRLF result lines normalized. The slash menu counts only the commands still hidden below its window and names an alias with its command on one row, built-in code skills are back in the palette,/resume opens as a prompt-region popup, overlay arrow keys stop scrolling the transcript, Cmd and Opt prompt deletion chords work, and a by-design static scene is labeled static rather than an offline bridge.CLI. A headless one-shot flushes its durable event log tail before exiting and prints the answer in one write instead of streaming it, and a headless run that refuses an interactive request now warns. The CLI exits quietly on a closed stdout pipe and keeps feature boot diagnostics inside the TUI. Update checking honors the output mode in ori update --check, restates its cache after an install, and tells source checkouts the truth while letting ori update repair missing release metadata. ori eval resolves auth through the shared login gate, harness launches use the ori login key rather than a project dotenv key, and the bundled ori SDK is injected so ori eval works in a fresh repo. The workspace .ori directory ignores itself, and the remote feature cache is reused when the ref has not moved.Eval. A guessed model slug is rejected at authoring time, the scoreboard stops reading “passed” when the test failed, and a baseline that can never exist is reported as such.Harnesses. A mid-turn 402 is clamped instead of losing the whole run, and the exhausted-402 error points at OpenRouter’s own remediation URL. Pi waits for its retry verdict before failing a turn, recovers the session after the agent process dies, and emits Anthropic cache_control breakpoints for ~anthropic slugs. Claude pins the base URL in the missing-key case. Peer exits and protocol errors project as terminal failures. A mid-session model switch applies to the running harness, and a terminal turn failure carries the usage accumulated so far.Runtime. A slow model stream is no longer severed at ten seconds, and a failed turn no longer emits a phantom model-less run.started. External skills are discovered up to the project root, built-in code skills are injected for plain ori runtimes, and a skill whose source directory is gone is skipped rather than failing the turn. ori harness test gets a per-case deadline, and runtime errors are standardized end to end.Telemetry. Telemetry stops reporting user paths and attributes agent runs by harness and surface.This release launches agents through
ori claude, ori codex, ori opencode, and ori hermes. It turns ori eval into a harness with judges, baselines, history, and reports. Docker agents now run behind a key-proxy with declared feature secrets, while agents can ask questions during a turn and the chat TUI gains broader support.Agent launches
- External harnesses. The CLI adds
ori claude,ori codex,ori opencode, andori hermes, which launch each harness through OpenRouter. - Coding persona. A bare
orilaunchesori code. Missing harness binaries prompt for installation. Fable registers as a first-class 1M-context model, and prompt economy works on the OpenRouter path.
Evals
- Judge criteria.
ori evalships starter judge criteria. Runs default to claude-opus-latest with a sonnet-latest judge, and the judge later pins to claude-opus-latest. - Reports and history. Each completed run reports its outcome and usage.
.ori/eval/history.jsonlstores run summaries, and each run can write a shareable markdown report. - Baselines and authoring.
ori evalasserts cost and duration ceilings and compares runs with a known-good baseline. A writing-evals skill joinsori code, selects a model surface, and uses the widened OpenRouter model catalog. The runner offers a consented Bun install when Bun is missing.
Chat TUI
- Themes. ^P or
/themeopens the theme picker. - Rendering. Streamed Markdown renders natively. A unified diff renderer lands. Streamed reasoning uses collapsible Thinking…/Thought for Ns rows. Links show an underline, and the workspace header row has a contrasting background.
- Tool calls. Consecutive calls collapse into ^O-foldable groups with per-tool expansion and hover highlights. Working indicators and an explicit expand/collapse affordance improve tool progress visibility. Live tool progress surfaces as it happens, and provider retry status also surfaces during calls.
- Input and sessions. A
?hotkey opens keyboard shortcuts. The first ctrl+c clears the draft, and a second press exits./quitaliases/exit. Resume sessions use their latest prompt as the name.
Docker deploy and the key-proxy
- Key-free agents. The agent container sends requests through a key-proxy that holds the credential. Container egress is locked to that proxy.
- Deployment image. A repo-buildable prod-emulation container ships with
dev:docker. The key moves to a secret file, the workspace becomes configurable, and the updater syncs managed deployment artifacts.
Feature secrets and interactive requests
- Declared secrets. Features declare needed secrets with a placeholder sentinel. The key-proxy substitutes declared secrets on the request leg, so the agent never sees them.
- Interactive requests. Agents ask questions during a turn. The TUI renders dialogs, and Pi gains
ask_user.
Workspaces, skills, and the CLI
- Workspace lifecycle. The global workspace gains an archive primitive and
ori workspace reset. - Skills.
ori skillsfetches built-in skills. External agent skills load from the launch cwd as a builtin, and theori codeagent gains a schedule skill. - CLI controls. An initial prompt accepts
--prompt,-p, or--prompt-file.ori coderuns an interactive update check with an early-access opt-in, andori --versionshows update status. Reasoning appears in logs, andmcp.jsonconfigures MCP servers for harnesses andfeature.ts.
Feature authoring
- Feature APIs.
FeatureApistyping is auto-generated, and per-pairuse()gating is dropped. Theuse("slack")provider type is available in the ori SDK. - Telemetry. An opt-in OTLP event exporter lands. Eval runs and activation latency are reported, and the installer tracks its telemetry.
Bug fixes
- Transcript rendering. The reasoning disclosure caret aligns with the row gutter. The resolved theme background applies to the terminal surface. Hover and expand UX unifies across transcript disclosures. The TUI stops tearing when the terminal is resized.
- Resize handling and footer status. The footer decodes percent-encoded worktree basenames. It shows the model a
~alias resolved to. It separates workspace from usage status. - Input and clipboard behavior.
/compactechoes the typed line instead of a synthetic prompt. Drafts clear properly and slash menus are capped. Clipboard images read without pngpaste. The clipboard image hint stops appearing for text copies. - Session startup and diagnostics. The first message renders when the runtime re-mints the session id. Remembered-model startup logs are silenced. Capture proxy logs stay out of the TUI. Apple Terminal prompt marks are avoided.
- Session reattachment. Sessions reattach so replies stop repeating.
- Pi usage. Pi surfaces token usage and cost, and pi turns are priced from OpenRouter’s streamed usage.
- Pi runtime and tool interactions. Pi allows custom ask-user responses. Its install and runtime use ori-as-bun, so fresh machines need no external bun.
- Pi elicitation UI. Pi
ask_userrenders a real selector without diagnostics corrupting the screen. It submits and clears its responses properly. - Claude result handling. A Claude run that exits without a result now fails properly. Bundled skills are disabled for
ori code. Anthropic model-tier env vars default on the OpenRouter path. - Codex authentication. Codex uses command auth so model metadata resolves.
ori claude/codex/opencode/hermeslaunch again. - Eval runner setup. The runner accepts the eval target positionally. It applies a 120s default per-test timeout.
- Judge validation and verdict parsing. The judge’s 0..1 score bound is enforced on every transport. It survives harnesses that ignore outputSchema and recovers verdicts from assistant-message envelopes.
- Bakeoff results and reports. Judge runs are told apart from bakeoff candidates. A cut-off bakeoff candidate stays in the results. The baseline scopes to the eval files that actually ran. The judge’s reason is quoted in the shareable report.
- Eval output and SDK checks. Dev progress notices stay off stdout in json mode. A scaffolded workspace typechecks against the shipped SDK.
- Writing-evals scaffolding. The writing-evals skill broadens its dispatch to bare model-selection prompts. It produces
*.eval.tsin any-language repos. - Telemetry and JSON output. The telemetry first-run notice names CLI commands and is boxed. Runtime diagnostics stay off the
--jsonenvelope stream.ori init --jsonis parseable and stops guessing the next command.ori initretries transient template download failures. - Workspace recovery and updates. A stale global workspace repairs itself so the dashboard feature builds on boot. Auto-scaffold fails loudly on a broken bun install. The update channel resolves from the installed release. Usage cost reports in alpha builds.
- Managed skills and workspace loading. Managed skills route through the docker key-proxy. Global workspace skills load in
ori code. Skills stop materializing into the launch repo. - Model aliases and telemetry details. The
-latestmodel aliases are tilde-prefixed.web_searchpoints at a configurable OpenRouter base URL. Telemetry records only the resolved command path, tags environment kind, and reports real CLI errors. - Runtime recovery. The runtime boots through a broken feature instead of dying. NDJSON turns stay alive with a 30s keepalive line. Host memory no longer grows without bound during long runs. A project feature can replace a built-in of the same name.
- Compiled feature resolution. Compiled binaries load SDK features and feature schedules from packages on disk.
- Schedule records. The schedule feature decodes persisted run records instead of coercing unknown status to ok.
- Docker and Slack. Every dev container gets its own telemetry install id. Slack persists queued turns and replays them after a restart. It reports an error when a turn dies without a terminal event.
This release adds typed hooks for feature authors, context rollover and compaction, and
/resume, /compact, and /context in the chat TUI. Slack gains a cancel command and early status, and the CLI adds an events panel to ori dev.Feature authoring
- Typed hooks. Features can register typed user-space hooks.
use()everywhere. Command, schedule, and chat contexts gainuse().
Context management
- Rollover and compaction. Sessions roll over at the context threshold. pi compaction uses the model context window, and native harness compaction is preferred at the threshold.
Chat TUI
/resumeand/compact./resumepicks up a previous session inside the chat TUI, and/compactroutes per harness.- Context visibility. The
/contextcommand and context window visibility land in theori codeTUI. - Status line and footer. The TUI shows the current git branch beside cwd and removes the redundant footer cache chip.
- Rendering and input. Tool output gains syntax highlighting and hover-brightened traces. Clipboard images paste with cmd+v/ctrl+v, and resumed transcripts warn on a harness switch.
Slack
- Cancel and status. A text-based “cancel” command stops stuck runs. An early loading status posts before turns enter the queue.
- Context and filtering. Undirected channel messages are filtered out. The per-turn Slack context shows the active harness and model.
CLI
ori devand diagnostics. An interactive events panel lands inori dev, which logs its resolved execPath/main/argv before dev and update restarts.- Remote features. Remote-features fetch retries over SSH when HTTPS fails.
- Coding persona. The coding persona gains dedicated code-practice skills.
- Eval.
ori evalmodel comparison uses OpenRouter’s live catalog instead of a hardcoded list. - Skills. The
openrouter-skill-slugmanaged skill pointer ships.
Bug fixes
- Chat TUI. The transcript is padded so the scrollbar and chatbox stop crowding content. Pickers now preselect the active harness and model. Copy falls back to
pbcopyon Apple Terminal. - Slack. The thread status pill clears once the reply bubble posts. Markdown tables auto-convert to native Slack table blocks. An unconfigured chat surface is skipped instead of failing.
- Logging, panel layout, and feature deps.
.ori/logsboundsraw.payloadand NDJSON line size. The split panel gets exact pane widths and full chat mouse forwarding. Composing--featuressources now installs the real workspace’s own deps. ori lint.ori lintparses.tsxfeature files, formats after fixing, and stops flagginguse()-only cross-feature deps as unused.- Host configuration and workspace paths. An explicit
--host+--portpair now readsfeaturesRootfrom the descriptor. An optionalorifile:dependency is recognized as an existing workspace. The dev/update re-exec argv no longer leaks the/$bunfspath. ASKILL.mdmodel no longer overrides the workspaceori.mdmodel. - Harnesses. Claude Code stops prefixing prompts with the stdin dash placeholder. pi runtime siblings are pinned to prevent an oauth drift crash.
- Runloop.
ori code’s harness cwd anchors to the launch directory instead of the global workspace. Managed skill links self-heal, with softer pointer fallback diagnostics.
This release adds the
command contribution kind, a deterministic /name action for the TUI and Slack. Workspaces can declare feature sources, while sessions, model picks, and tool rendering improve.Command contribution kind
/nameaction. A seventh capability authored as plain TypeScript dispatches pre-agent for people in the TUI and Slack, including DMs.
Declarative features
ori.mdfrontmatter. A workspace commits feature sources inori.md, while the repeatable--featuresflag supplies them per invocation.
Chat TUI
- Session switching. Sessions get a dedicated page with per-session transcripts, and session selection is smoother.
- Model and harness.
/modeland/harnesspicks are scoped. - Rendering. Tool call rendering is lighter, and the slash menu suggests feature commands and skills.
Feature authoring and deploy
- User attribution.
agent.invokecarries an optionaluserIdthrough the harness to durable session metadata. - Deploy. Headless deploy artifacts and a runbook ship, with a tunable resource backstop.
Bug fixes
- Slack. Prior-turn context now reads the full reply-cache text instead of Slack’s 200-char notification fallback. An expired trigger falls back to a thread reply while keeping the full reply modal-only.
- Terminal input and boot defaults. Raw wheel reports forward to the chat pane instead of being read as arrow keys. Boot failures surface instead of hiding behind a silent model fallback. The default model scaffolds to
openrouter/auto. - Remote feature validation. A remote feature root’s declared deps install after fetch. Feature directory names are validated at load time.
- Chat startup and turn routing. Chat starts faster. Streamed turns route per session and the footer is more compact. The
/modelpicker sorts by weekly popularity. - Model state persistence.
/modelslugs are validated. The last/modeland/harnesspicks are remembered across restarts. - Terminal compatibility. Bracketed paste stays enabled on Apple Terminal so multi-line pastes stay one prompt. Any-motion mouse mode and the cursor-color OSC are stripped from Apple Terminal host output.
- Harness.
IS_SANDBOX=1is injected into the claude subprocess env. Harness process subtrees are reaped on teardown instead of leaking. - Runloop. Invoke-stream failures end the turn instead of hanging.
Slack becomes a complete chat surface with cancel and recovery flows. The Remote features and the
api contribution kind and remote feature sources expand the framework. The CLI gains telemetry and update controls, while the TUI adds session commands, footer stats, and markdown rendering.Slack chat surface
- Cancel and recovery. A Cancel button sits beside See details and stops an in-flight turn. Its styling and size match neighboring controls. Interrupted tool rounds are rescued and reprompted instead of leaving a bare “done” marker.
- Dispatch and participants. Five bugs in dispatch, thread history, and finalize are fixed. Auto-mute no longer counts the agent as a participant and no longer misses other apps in the room.
- Response delivery. Successful runs return the real result instead of the See details placeholder. Prose recovery keeps prompts and tool output out of replies, and the loading indicator appears sooner.
- Modals and prompts. The expand modal opens with the ack. Expired triggers fall back to a thread reply, and the turn prompt is more insistent about always ending with a summary.
Remote features and the api contribution kind
- Remote sources.
ori start --featuresaccepts a remote GitHub repo path as well as a local directory. A GitHub App token fetches it, and multiple--featuressources collapse into one root. - API contribution. The
apicontribution kind ships with exports and daemon HTTP routes for feature APIs.
Chat TUI
- New session.
/new, aliased/clear, starts a fresh session without a process restart. - Footer and stats. The footer uses one line and a reserved toast row. It shows session cost, turn count, cwd, and the harness default model on the first frame.
- Model state. The resolved model reappears after a brief gap, and footer padding matches the TUI rhythm.
- Rendering and controls.
ori codehighlights GFM markdown. The slash menu supports arrow-key cycling, image attachment tokens are compact,/modeland/harnesswork mid-turn, and a second Ctrl+C is required to exit. Clicking a message restores prompt focus. - Terminal compatibility. Transcript links open on Apple Terminal. Stray bracket sequences and synchronized-output mode-sets are sanitized in the chat and schedules views. Mid-turn messages show transcript feedback.
CLI and telemetry
- Session resume.
--resumeand--sessionletori codereattach to a previous session. - Auto-update.
ori startsupports--alphaand opt-in--watchhot reload.ori updateshows progress during transient retries and skips downloads for the latest release. - Telemetry. Usage telemetry tracks installs and command usage.
- Auth and lint. OAuth-provisioned keys use the
key_labelvalue “Ori”.ori lintincludes markdown and folds in syncpack and knip. - Skills. Managed skills resolve from a frontmatter pointer at boot instead of a hardcoded path.
pi harness
- Prompt delivery. Large pi prompts pass through files instead of argv to avoid E2BIG.
- Reliability.
max_tokensis clamped, OpenRouter 402s retry, and native text selection inori codeis no longer swallowed. - Model ids. The model picker stores plain catalog ids, while the pi harness owns the
openrouter/prefix. Harness defaults also show plain catalog ids. - Attribution. OpenRouter attribution uses
https://or.bot.
Runtime
- Runloop. Headless runs gain a durable event log, workspace-anchored state, opt-in schedule jitter, and an overlap policy.
- State and performance.
ori/stateships in the SDK, and Slack uses the same store. Chat startup latency drops.
The chat TUI becomes an interactive surface with model and harness pickers, run steering, prompt recall, and rendered markdown. Slack, durable session logs, and the pi runtime also land. Read the breaking change before consuming runtime events.
Breaking changes
- Event tags. Harness runtime events no longer carry a boolean success flag. Each event uses a succeeded or failed tag for discrimination.
Chat TUI
- Model picker.
/modelswitches the OpenRouter model in the TUI. The list filters text-in/text-out models, supports:nitroand:floorvariants, and matches theori code--modelflag. - Harness picker.
/harnessswaps the active harness. - Run steering. Enter steers an in-flight run, and ESC interrupts it while keeping prompt focus.
- Prompt recall. Up/down recalls prompts, and
/exitquits the chat. - Rendering. Replies render as markdown, and URLs are clickable. Tool calls always show collapsible result boxes. The footer shows live token usage, cost, and a copy-on-select toast.
- Layout. Ctrl+R and
/refreshredraw on demand. Terminal focus and SIGCONT also trigger redraws. - Syntax highlighting. Syntax highlighting works in compiled binaries. Picker rows remain readable.
- Slash commands and footer. Enter runs the highlighted slash command on a half-typed prefix. The footer shows the current model once.
- Terminal layout. The TUI reflows on terminal resize. The input bar is wider and always shows the default model and harness. Ori strips stray bracket sequences on Apple Terminal.
Slack chat surface
- Slack. A built-in Slack chat surface lets an intern hold a conversation in Slack as in the TUI.
Concurrent sessions and durable logs
- Spawn-thread.
/spawnshows a usage hint, seeds concurrent sibling sessions from a shared parent, and keeps the picked model. - Durable logs. Every daemon mode writes a durable session log with metadata sidecars, so a run’s transcript survives past the process.
ori codepersona.ori codeboots against the global workspace with a built-in coding persona.- Alpha channel. An opt-in alpha pre-release update channel is available for early builds.
pi harness runtime
- Routing defaults. pi defaults to
z-ai/glm-5.2, forces the OpenRouter provider prefix, and routes unset sessions through OpenRouter. - Bun runtime. pi runs on the Bun runtime by default with graceful fallbacks, and pi is installed into an Ori-owned runtime directory rather than
bun add -g. - web_fetch SSRF. DNS resolves before the SSRF check, and the socket pins to the vetted IP to close the TOCTOU gap.
Reliability
- Process teardown. Harness timeout escalation to SIGKILL and stderr joins are bounded.
- Tool results.
ori coderenders Claude tool results. - Runloop and reload. Dev reload is atomic.
ori codeno longer creates a.agents/skills/feature-developmentsymlink.
This consolidation release adds feature compatibility and migration reporting. It also adds framework-owned linting, an evaluation harness, and scheduler reliability improvements.
Feature compatibility and migration
- Version provenance.
ori initstamps the generating CLI version intoori.mdfrontmatter. - Capability lifecycle. A bundled registry keyed by CLI version records capability introduction, deprecation, removal, and breaking shape changes. Its migration report uses the
scheduledefault-export to namedexport const scheduleas the canonical example. - Compatibility reports. Ori explains compatibility failures instead of letting older workspaces fail silently, and classifies findings as blockers, warnings, or notes.
- Surfaced reports.
ori features validateshows the report.ori updateexits non-zero on a blocker and advances the workspace baseline after a clean report. Existing-target sync inori init .also shows it. - SDK surface. Built-in and external feature imports converge on the bare
oriSDK specifier.
ori lint
- Framework rules.
ori lintships as a first-class command, and scaffolded workspaces no longer carry their own biome config.
Evaluation harness
- Author surface. The
ori/evalauthor surface andori evalcommand let features ship evaluations beside their code.
Scheduler
- Catch-up. Schedules can opt in to missed runs after restart instead of skipping them.
- Disable without deletion. A
disabledfrontmatter flag holds a schedule off cron while keeping it in the workspace.
Chat TUI and operations
- Tool-call rendering. The chat TUI renders tool calls with selectable verbosity levels.
- Auto-update.
ori startauto-update no longer waits on a Slack approval step.
Reliability
- Structured diagnostics. CLI errors keep their structured diagnostics instead of flattening them to strings.
- Daemon errors. The daemon serializes HTTP errors as a structured envelope.
The first tagged release brings up Ori’s declarative agent-building stack. It introduces feature contributions, a local runtime, the
ori CLI and dev loop, pluggable harnesses, and compiled binaries.Features and runtime
- Feature contract.
feature.tsuses the compiler-enforcedsatisfies FeatureModulecontract. Features contribute skills, prompts, generations, andgetModel, and agents run from feature boot results. - Local runtime. Resolved features build and boot locally. The runtime provides a built-in cron scheduler and the
schedulecapability, provider selection, multiple schedules per feature, a SQLite state store, and edit-mode reload.
CLI and interfaces
- Commands. The CLI includes
init,dev,start,code,login,update, andversion.ori startruns a headless bot server,coderuns a local coding agent, andloginuses OpenRouter PKCE.ori initscaffolds from a default template and rootori.mdpersona, writes anAGENTS.mdguide, and validates the workspace. - Chat and terminal UI. The chat TUI supports
Shift+Enter, multi-line paste, OSC 52 drag-to-copy, and a blinking caret. The split-paneori devlayout supports mouse resize and chat scrolling. - Harnesses and tools. Pluggable harnesses include PI
web_searchandweb_fetch. The release includes a local Slack bridge.
Operations and packaging
- Logging. Features log through
ori/logger, and structured records reachori logs. - Packaging. Compiled CLI binaries ship with
install.sh. Version-matched docs ship into.ori/docswith agent pointers. - Operations.
ori startsupports auto-update with a severity threshold and interactive Slack approval. OpenRouter attribution uses injected headers with actionable authentication hints when authentication fails. An agent-facing machine output mode supports scripting.