Brennan BensonandClaude Opus 5.5 d1d4ac3714 Stop retrying final native agent startup refusals (#37213)
* Stop retrying final native agent startup refusals

* Retire startup opens after exit and keep final answers across exit orders

- A start flight cancels its outstanding provider open on every ended path; a direct attach keeps the child-bound abort.
- The post-exit answer join is bounded by the remaining startup ceiling.
- An elapsed initialize deadline keeps its words when the exit wins the race.
- Codex cloud config temporary declarations retry on thread/resume as on thread/start.
- A Claude exit before acquisition publishes waits briefly for initialize's answer, so a rejected initialize stays final.
- The Claude startup stdout observer parses only failed root results, within the shared line budget, and detaches once it has one.

* Keep saving the last turn when a Claude process exits after startup

The exit path skipped saving the resume cursor whenever Orca had not
dispatched a message, but turns can settle without an Orca dispatch and an
already-started idle session still owns its exit save. Only a start that
died before initialize answered and before any turn now skips the save and
gets the short cleanup bound. A failed result captured before the first send
counts as a startup failure only when no turn ran in that process.

* Merge duplicate test imports

* Bound resumed Claude starts that exit during catch-up and keep untyped refusals final

Exit-time cursor writes and the short close bound now key on the host-awaited
open actually returning, so a resume that exits while transcript catch-up is
pending publishes ended at once instead of waiting on a cursor write. The
startup observer no longer reads an answer from an oversized line's prefix.
A typed acquisition refusal without a dedicated failure kind keeps its
refusal marker, so an exit-only start stays final.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test: pin the fake clock in the resumed-exit start test

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test: bound startup answer parsing by candidates at N and 10N records

Drive the startup observer through the shared framer at 100 and 1000
irrelevant hook and assistant records, and assert that only the failed
root result is parsed, one capture happens, the data listener retires,
and no parse follows capture or finish, with identical counts at both
populations.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test: shape startup fixtures like the real providers

Held opens in the launch test now reject when their child exits, as a
real connection does, and the temporary open failures use the provider's
actual overload envelope. Also fixes the test-only type errors CI found.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test: use the real Codex canonical-filename refusal wire capture

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(claude): align merged startup tests with answer-wins and final generic results

An exit with no stderr leaves the start to Claude's buffered result, so that result is the start's one failure. A generic startup failure result is final after one start; only an unanswered exit retries, and the recovery case now recovers from such an exit.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(claude): let a final initialize answer outrank a physical exit reported fault-first

The connection reports one physical root exit through onFault and then
onExit. It now marks that error as an observed exit before either
callback, and acquisition's onFault leaves a proven exit to onExit's
bounded initialize join, so a final initialize rejection that arrives
just after the exit owns the start instead of a retryable bare exit.
Typed refusals, live-root faults and processless spawn faults still end
the start at once.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(claude): await a macrotask instead of real time in the exit-answer start test

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(claude): replace casts in the exit-answer start test with an annotation and guard

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(agent-session): let each adapter decide a failed start, and give the start one owner

A start that the agent refused could still be replayed: the host joined a
child's exit with its open's error and rebuilt the meaning from provider error
classes, and an answer written just before exit could lose to the exit itself.

Each adapter now answers its own acquire, open and required startup reads with
one typed result (final, or safe to run once more), keeping a final answer
beneath the exit proofs cleanup adds. Transports keep reading answers already
written after the agent stops reading or exits, within the existing one-second
drain, and only a drained exit of the agent's own process counts as unanswered.

The child's open is the single owner of the start's outcome: it awaits the
adapter's result and required settlement, records the verdict once, writes a
direct attach's failure row, and stops the child. The start flight reads that
verdict instead of joining exit and open. A start whose adapter gives no outcome
within five seconds of its own process exit fails through a timer that no
bookkeeping can hold, and the original start deadline stays in force until the
start is decided.

Tracked in STA-12982.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(agent-session): bound a failed start's lifetime and keep Orca's own refusals final

- A start's ending is kept only while its owner can read it: admitted when the
  acquisition starts, retired when the owner finishes, the acquisition fails, or
  the session is released or closed. An exit after ready records nothing.
- Codex takes its own exit as the start's wake from the moment the process is
  seen, before spawn bookkeeping, and a failed acquire keeps that exit as cause.
- Once the handle or rewind write began, an unanswered death no longer decides
  the start: the write's refusal stays final, and the exit grace bounds the wait.
- Stop, the deadline and the exit grace release the start's callers at once; the
  failure row is queued first but never awaited by them.
- An ACP connection Orca closed before the exit keeps that cause, so cleanup's
  exit is not read as the agent ending unanswered.
- A request deadline keeps its words when an exit is seen around it.
- A direct attach whose exit settled first keeps the exit's words.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(agent-session): let a start's owner, not its live route, keep and retire its ending

Exit acknowledgment drops a session's route while the start's owner may
still be reading how its acquisition ended, so the router now reaches the
adapters by exact generation for both the read and the retirement. Claude
admits its acquisition before publication can observe an exit.

An ACP start Orca refused before the agent exited now fails with that
original refusal as the adapter's final verdict, through every cleanup
wrapper, instead of a generic closed-while-starting error.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(agent-session): give the worker retry fake the Codex adapter's typed start verdict

The host now reads only the adapter's verdict on a failed start, so the
fake open rejects with the verdict the Codex adapter returns for its
declared overload instead of the raw request error.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(agent-session): let Claude's own exit record decide a start the SDK reports as exited

A Claude CLI missing from the launch PATH makes the supervisor exit 127 with the spawn error on
stderr. The SDK rejects initialize with only "process exited with code 127", and the start took
that report as its cause, so the missing CLI was lost: the chat said Claude stopped before it
finished starting and the start was retried as an unanswered exit. The SDK's exit report is no
answer; the start now reads Orca's own record of that exit, which carries the missing CLI.

Also brings two ACP tests in line with the typed start verdict (a signed-out open rejects with
the adapter's notSignedIn verdict) and the bounded drain after the agent stops reading.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* perf(agent-session): stop parsing Claude's startup output once its failure is captured, and pin start and drain work at N and 10N

A failed root result captured mid-chunk still let the rest of that chunk be framed and parsed;
nothing reads those records once the start has its answer. The observer now accepts no further
line once it has captured the failure, and drops what that chunk left in its framing state.

Two contract tests drive the production start owner and transports at N and 10N:
- starts through the host, router and Codex adapter (ready, final refusal, safe replay, deadline
  expiry): each start's owner, abort handles, tracked attempt, flight, buffered exits, adapter
  receipts, root-exit subscription and deadline timer return to baseline, a waiting start holds
  a fixed set, and spawns, opens, router reads and retirements, timers and failure rows grow
  exactly linearly;
- ACP, JSON-lines and Codex connections with a pending call, answered at exit or abandoned after
  the agent stops reading (with duplicate terminal signals): each call settles once, each output
  drain finishes once and leaves no listener or timer;
- an ACP start refused at sign-in keeps its own first cause and leaves no receipt or child.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(agent-session): decide a failed start once from everything its adapter saw

A failed start was decided from whichever error reached the decision first, so a later or
poorer report could replace the one that explained it: the SDK's exit report hid the missing
CLI Orca's own exit record carried, and an output drain let a load deadline give way to an
unanswered exit. One shared rule now weighs every observation an adapter kept. The start may
run again only if every observation allows it, and the most specific one explains it: what the
user can act on, then a request deadline, then the agent's answer, then Orca's record of its
exit, then an SDK's report of that exit. Claude, Codex and Pi pass their exit records alongside
the operation's error instead of choosing between them.

An ACP agent that exited while its output drained no longer gets a replacement session after
its load fails. JSON-lines output that reports both its end and its close now starts one wait,
so the agent's exit can always cancel it.

Contracts at N and 10N: JSON-lines output ending while its root lives, real Claude and Pi
adapters retiring every start's receipt, and a user Stop whose cleanup is held then released.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(acp): keep a missed saved-session load deadline instead of opening a replacement

A live agent that let session/load run past its deadline fell through to the
takeover path, and the replacement session/new answer then replaced the
deadline the user had to see. The takeover predicate now propagates any
provider request deadline unchanged, read through the existing typed timeout.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-11-03 10:53:21 -08:00
2026-03-16 22:27:51 -07:00
2026-03-28 10:19:14 -07:00

Orca Orca

GitHub stars Total downloads across all releases License: MIT Join the Orca Discord Follow Orca on X Supported platforms: macOS, Windows, and Linux

中文 · 日本語 · 한국어 · Español · Français · Português · Türkçe · Русский

The AI Orchestrator for 100x builders.
Run Codex, ClaudeCode, OpenCode or Pi side-by-side — each in its own worktree, tracked in one place.

Download Orca

Orca desktop app running agents in parallel worktrees, with the Orca mobile companion app in the corner

Features

Mobile Companion

Monitor and steer your agents from your phone — get notified when an agent finishes and send follow-ups from anywhere.

iOS App Store · Android APK · Docs →

Orca desktop with the mobile companion app

Parallel Worktrees

Fan one prompt across five agents, each in its own isolated git worktree — compare the results and merge the winner.

Docs →

Parallel worktree orchestration

Terminal Splits

Ghostty-class terminals with WebGL rendering, infinite splits, and scrollback that survives restarts.

Docs →

Terminal splits

Design Mode

Click any UI element in a real Chromium window to send its HTML, CSS, and a cropped screenshot straight into your agent's prompt.

Docs →

Embedded browser and Design Mode

GitHub & Linear, Native

Browse PRs, issues, and project boards in-app — open a worktree from any task and review without a context switch.

Docs →

GitHub and Linear task workflows in Orca

SSH Worktrees

Run agents on a beefy remote box with full file editing, git, and terminals — auto-reconnect and port forwarding included.

Docs →

Remote worktrees over SSH

Annotate AI Diffs

Drop comments on any diff line and ship them back to the agent — review, edit, and commit without leaving Orca.

Docs →

Annotate AI-generated diffs

Drag Files to Agents

VS Code's editor with autosave everywhere — drag files or images straight into an agent prompt.

Docs →

Drag files and images into an agent prompt

Orca CLI

Agents drive Orca too — script every workflow with orca worktree create, snapshot, click, and fill.

Docs →

Script Orca from the CLI

Also in the box:

  • Quick open — Search across worktrees, files, agents, commands, and repo context without leaving your flow.
  • Account switcher & usage tracking — See Claude and Codex usage and rate-limit resets, and hot-swap accounts without re-logging in.
  • Rich repo previews — Preview Markdown, images, PDFs, and repo docs in the workspace.
  • Computer Use — Let agents operate desktop apps and visible UI when a workflow needs real interaction.
  • Notifications and unread state — Know when an agent finishes or needs attention, then mark threads unread to come back later.
  • And many, many more — we ship daily, so this list is perpetually behind. The changelog is the real feature list.

Supported Agents

Works with any CLI agent — if it runs in a terminal, it runs in Orca.

Claude Code logo Claude Code   Codex logo Codex   Grok logo Grok   Cursor logo Cursor   GitHub Copilot logo GitHub Copilot   Muse logo Muse   DeepSeek Harness logo DeepSeek Harness   ZCode logo ZCode   OpenCode logo OpenCode   MiMo Code logo MiMo Code   Amp logo Amp   OpenClaude logo OpenClaude   Antigravity logo Antigravity   Pi logo Pi   oh-my-pi logo oh-my-pi   Hermes Agent logo Hermes Agent   Devin logo Devin   Goose logo Goose   Auggie logo Auggie   Autohand Code logo Autohand Code   Charm logo Charm   Cline logo Cline   CodeBuddy logo CodeBuddy   Codebuff logo Codebuff   Freebuff logo Freebuff   Command Code logo Command Code   Continue logo Continue   Droid logo Droid   Kilocode logo Kilocode   Kimi logo Kimi   Kiro logo Kiro   Mistral Vibe logo Mistral Vibe   Qwen Code logo Qwen Code   Rovo Dev logo Rovo Dev   + any CLI agent


Install

Desktop — macOS, Windows, Linux

Or via a package manager:

# macOS (Homebrew)
# Always use the full name; bare `orca` is an unrelated Plotly cask.
brew install --cask stablyai/orca/orca
# Upgrade an existing install
brew upgrade --cask stablyai/orca/orca

# Arch Linux (AUR) — or stably-orca-git to build from source
yay -S stably-orca-bin

Mobile Companion — iOS, Android

Pair with your desktop app to monitor and steer your agents from your phone.


Community & Support

  • Discord: Join the community on Discord.

  • Twitter / X: Follow @orca_build for updates and announcements.

  • WeChat: Scan to join the Orca community WeChat group.

    WeChat QR code for the Orca community
  • Feedback & Ideas: We ship fast. Missing something? Request a new feature.

  • Privacy: See the privacy & telemetry docs for what anonymous usage data Orca collects and how to opt out.

  • Show Support: Star this repo to follow along with our daily ships.


Developing

Want to contribute or run locally? See our CONTRIBUTING.md guide.

The relay that pairs the mobile app with a desktop host is also in this repository under cloud/, with a separate pnpm workspace and setup guide.

Orca contributors

GitHub star history chart for stablyai/orca

Signed Builds

Windows code signing sponsored/provided by SignPath.io, certificate by SignPath Foundation.

License

Orca is free and open source under the MIT License.

S
Description
Orca is the ADE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
Readme MIT
2.8 GiB
Languages
TypeScript 95.4%
JavaScript 3.5%
Swift 0.3%
CSS 0.2%
Python 0.2%
Other 0.1%