mirror of
https://github.com/stablyai/orca.git
synced 2026-10-31 00:57:45 +00:00
Brennan Benson
7a78dbd3c7
* feat(native-chat): record which agent run wrote each journal item Each journal item row, subagent roster entry and background-task entry now records ownerFence: the lease fence of the agent run whose output produced it. Optional, persisted, no row version bump and no migration. The journal's itemFence() returns it, so the turn-verdict rules and per-run provider activity read the producing run rather than the writer's fence. * feat(native-chat): one current-work projection for every host reader; record death first The host decides what work is current in one place (structured-agent-session-current-work.ts) from the lease, the exit this host saw, the runtimes it replaced, its live child and each item's provenance. Every host reader (queue, Stop, prompt answers, commands, status feed, idle sweep, teardown snapshot, history pages and frames, orchestration mailbox) reads it, so a dead agent's leftovers stop reading Working or holding the queue even when their cleanup write failed. The journal's activeTurnId()/liveTurnScope() become the raw runningTurn(). On exit the host records the child's end first and signals generationEnded, which republishes frames, the status feed and settlement waits and wakes the queue drain with no journal row. Pages and frames carry optional working and actionablePromptIds fields. * feat(native-chat): clients read the host's working answer; refuse answers to a stopped agent plainly Desktop and mobile prefer the host's working and actionablePromptIds fields when present (merged across batches like latestTurn) and fall back to today's row derivation against an older host. A prompt an ended agent raised stops looking active; answering it is refused before any provider sees it, with "<Agent> has stopped, so your answer was not sent. Send a message to continue." Cancelling it only dismisses it in the journal: it no longer triggers the chat's Stop, pauses the queue or withdraws queued cards. * feat(native-chat): one settlement transaction for everything an ended run left An exit now settles, in one lifecycle batch planned at its own turn in the write queue, everything its run left (turn, calls, prompts, reasoning, unanswered sends) plus leftovers of earlier ended runs, and only the subagent roster entries and background tasks those runs observed, by each entry's own provenance. Main's stale settle moves to stale-state-settlement.ts with fence scoping; dead-generation-settlement.ts becomes a planner; leftover-settlement.ts combines them. Resume eligibility widens in one case: this host saw the root exit at the lease's fence and the release write failed; the acquisition replaces the owner only once a probe proves it gone. * refactor(native-chat): keep structuredAgentSessionHandsOver file-local until a caller needs it Nothing outside the projection calls it in this part; Part 5 exports it for the start-failure row. * test(native-chat): fewer provenance tests: fold reopen and unstated-owner checks into the main case Drops the epoch-replacement case (Part 2 rewind test reads the same values through the same rebuild) and the corrupt-value case (no version writes one). * test(native-chat): drop current-work mirrors: the viewKey unit case and the still-live rewind case The flow cases keep every outcome: dead leftovers read as nothing, Stop has nothing to stop, a waiting card drains, a live owner publishes as before, and a rewind never promotes dead work. * test(native-chat): fewer client tests: one flow per client, one fallback each Keeps desktop and phone Working from the host (with the older-host fallback), the reducer merge contract, the ended question Cancel that only dismisses, the agent-named refusal and the resolution receipt. Drops per-helper prompt tests, the transcript-placement test, the stop-note client reduction, the Steer cards that match main, and per-reason refusal rows. * test(native-chat): fewer settlement tests: keep the exit-delivery flows and the per-entry scoping flow Drops the no-added-delay timing check, the frame-once check, two stamping unit cases and the Codex combined case (the roster flow test covers its outcome). * test(native-chat): drop the resume-eligibility predicate mirror; move the unstamped-entry case to Part 1 * test(native-chat): an unstamped roster entry from an older build reads as its row provenance Moved here from Part 4 so it sits beside stampJournalEntryProvenance. * test(native-chat): build the older-build roster row inline, after main removed its body helper Main #27947 writes one row per subagent and deleted journal-subagent-group-body; a roster row from an older build is still what this case reads. * refactor(native-chat): provenance is per item only; drop per-entry stamping Main's #27947 writes one saved row per subagent, and each background task already has its own row, so an item's own ownerFence names the run of every entry a producer still writes. Per-entry stamps inside a row mattered only for the shared roster rows saved before #27947, and no run writes into those any more, so the row's own provenance settles them correctly. Removes the entry stamping, the optional ownerFence on roster entries and background tasks, and their tests. * fix(native-chat): an item's provenance is set when it is created; revisions keep it A newer run's provider can revise an item an older run created: Pi's start and an ACP session/load both report context usage onto the newest saved turn. Restamping that turn with the newer run let the newer run's exit settle the older run's unverifiable turn as interrupted. Only creation states an owner now, as main's itemFences rule did. * test(native-chat): keep the provenance test below the production lines it covers * fix(native-chat): an ended agent's card Cancel commits with its receipt, through dismissPrompt The ended-generation dismissal wrote the cancelled resolution without the Cancel's operation receipt and without the dismissal's failure handling, so a receipt write failing after it would answer "outcome unknown" over a card already closed. It now goes through dismissPrompt, which carries the uncommitted receipt in the same append and answers dismissalNotSaved on failure; a null `answer` skips the adapter, since no live provider holds an ended generation's card (Pi's adapter refuses a fence with no live session outright). * Native chat: a resumed subagent or Claude background task belongs to the run executing it A subagent row that run 1 created and run 2 resumed kept run 1's creation fence, so settling run 1's death marked run 2's live subagent unverifiable. Claude background tasks reuse their rows the same way (a matching parent tool use, or a parentless row). Rows of reusable child execution now also name their execution owner: the fence of the child process whose report the write copies, captured when that report arrives, never the lease that is current when a queued write flushes. - Subagent rows: the host captures the live child's fence when its child-work report changes the records, and rewrites a running subagent another child now reports even when its text is unchanged. - Claude background-task rows: written through that child's own event sink, which stamps its bound fence. - Creation provenance (`ownerFence`) is unchanged. Writes that do not state an owner (settlement, Pi's start usage, ACP session/load replay, ACP tasks) keep it, and a row without one falls back to its creation fence. `journal.itemFence()` now answers whose execution an item is (`executionOwner ?? ownerFence`), so every settlement and liveness read, inside its write transaction, follows the run that resumed the work. Also drops the event sink's compatibility alias for the identity resolver (the Codex sink names the resolver itself), keeping the sink under its line limit. * Native chat: a subagent row is written through the child whose report it copies Review ofc0add67cca: the subagent row writer read its execution owner from the host's installed child when a report landed. A child that is still being acquired reports a resumed subagent before the host installs it, so its row was written at once under the old owner, then the attach's settlement of what the previous child left marked it unverifiable, and nothing rewrote it until the next report. Each adapter (Claude, Codex, ACP) now hands its child-work report to the host with that child's own event sink, the write path its rows already take. The host passes it through the status feed to the subagent row writer, which writes the rows that report changes through that sink with `executes`. They land in the child's own order, after the attach's settlement, at the fence the sink is bound to, and that fence is the row's execution owner. A running subagent a child has not written yet is written once even when its text is unchanged. Changes the host makes itself (a Stop's acknowledgement, an end) still write directly and keep the row's owner. The shared child-work evidence type is unchanged: the sink travels as a separate argument. Test: a Claude helper the next child reports again while it is still starting stays working after that start's settlement, through the shipping adapter and host. It fails onc0add67cca(the row lands `unverifiable`). * Native chat: the subagent row writer always awaits a promise CI's type-aware lint flagged the row writer handing `Promise.all` a value that could be undefined when a sink has no `written()` barrier. A sink without one now reads as landed. * Native chat: keep refused subagent writes, and carry both owners through a rewind Review ofb161aa38b1/129f36d49b: - `admitChildWork`'s producer is optional again, so the host's own turn-started report compiles. - A subagent row written through the reporting child's sink uses the sink's checked admission. A write it refuses under backpressure is queued again once what the sink holds has landed; a refusal or a failed write frees that child's claim on the row, so a later report retries it. - A rewind carried `itemFence()` (now whose execution an item is) into the rebuilt row's creation `ownerFence`, and dropped `executionOwner`. Retained rows now carry both as stored (`journal.itemProvenance`), the rewind record keeps `executionOwner` beside `ownerFence`, and both rebuild paths write it back: epoch replacement and the context rewind, which builds its rows from the same replacement items. Test: the two-fence test now rebuilds the epoch from the resumed rows and settles the first run's death again; the resumed rows stay live. It fails if the epoch replacement drops `executionOwner`. * test: run the resumed-child owner test in the Node SQLite project It opens a real journal database, so it belongs on the list the SQLite runtime boundary test checks.
Orca
中文 · 日本語 · 한국어 · Español · Français · Português
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
Features
Mobile CompanionMonitor and steer your agents from your phone — get notified when an agent finishes and send follow-ups from anywhere. |
 |
Parallel WorktreesFan one prompt across five agents, each in its own isolated git worktree — compare the results and merge the winner. |
 |
Terminal SplitsGhostty-class terminals with WebGL rendering, infinite splits, and scrollback that survives restarts. |
 |
Design ModeClick any UI element in a real Chromium window to send its HTML, CSS, and a cropped screenshot straight into your agent's prompt. |
 |
GitHub & Linear, NativeBrowse PRs, issues, and project boards in-app — open a worktree from any task and review without a context switch. |
 |
SSH WorktreesRun agents on a beefy remote box with full file editing, git, and terminals — auto-reconnect and port forwarding included. |
 |
Annotate AI DiffsDrop comments on any diff line and ship them back to the agent — review, edit, and commit without leaving Orca. |
 |
Drag Files to AgentsVS Code's editor with autosave everywhere — drag files or images straight into an agent prompt. |
 |
Orca CLIAgents drive Orca too — script every workflow with |
 |
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
Codex
Grok
Cursor
GitHub Copilot
Muse
DeepSeek Harness
ZCode
OpenCode
MiMo Code
Amp
OpenClaude
Antigravity
Pi
oh-my-pi
Hermes Agent
Devin
Goose
Auggie
Autohand Code
Charm
Cline
CodeBuddy
Codebuff
Freebuff
Command Code
Continue
Droid
Kilocode
Kimi
Kiro
Mistral Vibe
Qwen Code
Rovo Dev
+ any CLI agent
Install
Desktop — macOS, Windows, Linux
- Download from onOrca.dev
- Or grab a build directly: macOS Apple Silicon · macOS Intel · Windows (.exe) · Linux AppImage · All builds
Or via a package manager:
# macOS (Homebrew)
brew install --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.
- iOS: Download on the App Store
- Android: Download APK · Install guide
Community & Support
-
Discord: Join the community on Discord.
-
Twitter / X: Follow @orca_build for updates and announcements.
-
WeChat: Group 11 is full. Scan to join the Orca community WeChat group 12.
-
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.
Signed Builds
Windows code signing sponored/provided by SignPath.io, certificate by SignPath Foundation.
License
Orca is free and open source under the MIT License.