Files
orca/AGENTS.md
T
Jinwoo HongandNeil 3f8d296b8d fix(macos): preserve terminal permissions across app updates with a small signed helper (#27021)
* fix(macos): preserve terminal privacy identity across app updates

* test(macos): make launchd fixtures portable across host platforms

* fix(macos): probe protected contents and close handles on denial

* fix(i18n): register terminal permission status labels

* fix(macos): fork the daemon when no stable-bundle job can claim the endpoint

A copy, signature, plist or deadline failure before bootstrap, a bootstrap
launchd refused with the job absent, a job whose daemon exited before
answering, and a foreign endpoint all threw out of the launcher. The app then
lost its persistent daemon and fell back to local PTYs, where main forks one.
These now fall through to the existing fork path, which also adopts a foreign
occupant normally instead of as a degraded endpoint. A timed-out bootstrap or
a job still running/unreadable stays fatal, so no second daemon races it.

* feat(telemetry): split daemon origin events by stable-copy launch method

daemon_adopted, daemon_pty_cwd_denied and daemon_pty_cwd_readable gain a
closed-enum launch_method (stable-copy / app-fork / unknown), read from the
spawner path the daemon recorded, so the update fix is measurable per daemon
even when an old forked daemon is adopted across updates. The reliability
gate is a test manifest with no runtime switch, so there is no gate state to
report; spawner_path_class reports the private copy as 'other' alongside
other non-/Applications installs.

* refactor(telemetry): report the stable copy as a spawner path class

The separate launch_method property duplicated what the recorded spawner
path already says. spawner_path_class gains `stable-copy`, classified from
the same path under the private runtime root, and launch_method, its enum
and classifier are removed from all three .strict() daemon origin events.
A deleted copy still reports `missing`.

* refactor(settings): gate terminal-host Full Disk Access on macOS once

The permissions handler already asks for the terminal-host status only on
macOS, so the status function's own platform check and the pane's
'unsupported' branch could never fire. The off-macOS guarantee is now
tested at the handler.

* fix(macos): keep the fork fallback inside the startup PTY gate

The stable launch could run until 55 s after the launch started, so a late
"unavailable" verdict left the fork fallback's own 10 s readiness wait to
finish after the 60 s gate had already moved the app to local PTYs.

The launcher now gives the stable attempt one handoff deadline: the gate
minus the fork's readiness timeout and the 5 s lease/adapter allowance
(45 s). Copy, signature check and plist conversion share one abort signal
that ends early enough for bootstrap and the job's own readiness wait to
meet that deadline; every command otherwise keeps a fixed timeout. An
unavailable verdict past the deadline is reported instead of forking a
daemon that cannot arrive in time. The per-step budget module is gone (its
deletion landed in the previous commit by a staging slip).

The launch body is now two blocks: preparation (plist, plutil, deadline
check) fails as unavailable after removing the copy and job file, and
bootstrap runs in its own try/finally that always deletes the job file.

Retirement is simpler and claims more orphans: one in-flight collection per
process instead of a per-root map and rotating 20-record scan, and the job
record (written before copying) no longer carries a bundle name because
lsof +D scans the whole runtime directory (measured equal cost, same
result on a live copy). A never-submitted copy is retired once its producer
has exited and its job is missing, so a bootstrap timeout or a crash
mid-copy no longer leaves a copy forever; the open-file check still guards
deletion. Collection now starts from the launcher after the attempt.

* docs(macos): ship the daemon launch rules, not the investigation

The signed investigation harness left the tree in the previous telemetry
commit by a staging slip; its results and pinned link stay in the PR. This
trims the reference doc to why the copy and launchd exist, the fallback
states and the cleanup rule, trims the reliability gate to what its
validator needs, and points AGENTS.md at the reference doc next to the
Windows relocation entry.

* fix(macos): hand off to the fork as soon as the stable job exits

A copied daemon refused at exec (for example by Gatekeeper) never answers,
so every fresh start paid the full 10 s readiness wait before the fork
fallback. The readiness loop now reads the job state every 500 ms and
abandons as soon as the job is stopped or missing; a running or
unreadable job is still waited out and never forked beside.

Also drops the dead forkEntryPath the stable launch never read (its
options type now omits it) and un-exports the handoff constant.

* fix(macos): keep daemon copies out of Spotlight and unregister retired ones

Each runtime copy is a full Orca.app, so Spotlight indexed it as an app
and LaunchServices kept a record for it that outlived the copy. A
registered copy claims orca: links and Markdown/CSV files, and becomes
their default handler whenever /Applications/Orca.app is not registered.

New copies go under runtime-*/app.noindex/, which Spotlight skips and which
stops the record LaunchServices made while the bundle was being copied.
Retirement now runs lsregister -u on every bundle in the copy (both the
app.noindex layout and the earlier runtime-*/Orca.app layout) once lsof
shows it unused, then deletes it. A running copy stays registered,
because Local Network access may be resolved through that record.

Collection and stable-copy telemetry match the runtime-* directory, so
copies made before this change are still adopted, classified and retired.

* docs(macos): record the LaunchServices trade-off and the rollback plan

The reference doc now explains why copies sit in a .noindex folder, why
only retired copies are unregistered, and how to turn the copy launch off:
make launchMacDaemonFromStableBundle return null, and keep the retirement
code until no copy-launched daemon can remain, because the old code never
deletes copies.

A test pins that rollback: with the stable launcher returning null, the
launcher forks and still runs the copy collection.

* fix(macos): keep daemon copies registered so Local Network can resolve them

The `.noindex` folder kept copies out of Spotlight and LaunchServices, but
Local Network access is resolved from the signing identifier through
LaunchServices to the executable's Mach-O UUID. After an update that changes
Electron only the copy carries the running daemon's UUID, so an unregistered
copy risks the same Local Network loss main has today. Copies go back to
`runtime-*/Orca.app`; retirement still unregisters each copy before deleting
it, so deleted copies leave no LaunchServices entry. The remaining trade-off:
while `/Applications/Orca.app` is missing, a running registered copy can be
picked for `orca:` links and documents.

* fix(macos): hide daemon copies from Spotlight and tie the job to Orca

Spotlight listed each runtime copy as a second Orca app. New copies go
back under runtime-*/app.noindex/Orca.app, which Spotlight skips.
Retirement and stable-copy telemetry match the runtime-* directory, so
copies in the earlier runtime-*/Orca.app layout still retire and classify
the same way.

Local Network access identifies the daemon by its code signature and
main-executable UUID, not by a LaunchServices record, so an unregistered
copy keeps it. The launchd job now sets AssociatedBundleIdentifiers to
com.stablyai.orca, which is how a job not installed through SMAppService
names the app it belongs to. Copies are never registered explicitly;
retirement still runs lsregister -u before deleting each one.

* feat(macos): package a signed terminal host helper that runs the daemon on the pinned Node

Contents/Helpers/Orca Terminal Host.app carries the pinned Node, the daemon's
out/main JS and this slice's node-pty, signed inside-out as com.stablyai.orca.
afterPack boots it through one PTY round trip, then fails the build on layout,
Info.plist, macOS floor (NODE_RUNTIME_DARWIN_MINIMUM_OS vs Node's minos),
architecture or signature drift.

* ci(macos): cache and prefetch the pinned Node before packaging

Packaging now fetches the darwin Node from nodejs.org, whose download has no
retry. Each mac workflow caches it with the electron-builder downloads, keyed on
the pin, and warms both darwin targets inside the existing retry wrapper.

* feat(macos): launch the daemon from a copy of the terminal host helper, not the whole app

The stable launch now clones Contents/Helpers/Orca Terminal Host.app (~124 MiB
instead of ~628 MiB) and runs its Node on its daemon entry, without
ELECTRON_RUN_AS_NODE. The copy must still match the running app's designated
requirement. A missing helper falls back to the fork, and macOS below the
helper's Node floor forks without copying.

* docs(macos): describe the terminal host helper, its floor and its identity

* fix(macos): bound the terminal host boot check by its deadline and always clean up

A daemon that stayed silent hung packaging past the 20 s deadline; one that closed the
connection left the hello unsettled so Node exited 0 without cleanup; a malformed line threw
uncaught. Every connect, hello and PTY wait now races the deadline, close/error/parse failures
reject, and the daemon and scratch dir are removed on every path.

* ci(macos): cache the pinned Node on its own key with no restore-keys fallback

Sharing the electron-builder cache let old Node archives accumulate through restore-keys.
Also moves release-mac-build's Electron cache comment back onto its step.

* fix(macos): reject a universal slice for the terminal host and pin its name in one test

packageMacTerminalHost turned an unknown arch into darwin-undefined; it now throws through the
shared mac arch mapping. A test asserts the packager and the desktop agree on the helper bundle
and executable names, replacing a Linux/Windows test that could not fail.

* fix(macos): pass the launchd adoption lease its options object after the main merge

* test(macos): accept a reset as the daemon closing the boot-check connection; keep only passing gate commands

* fix(daemon): bump the terminal daemon protocol to v45 so new macOS tabs start in the terminal host helper

* fix(macos): ship the daemon's required packages in the terminal host and boot it outside the build tree

* style(macos): use a template literal in the terminal host resolution check

* fix(macos): derive the terminal host's packages from the daemon's own module graph

* fix(macos): treat any non-literal daemon load as computed, follow directory loads, and keep nested dependencies

* refactor(macos): move terminal-host telemetry, Settings status and Helper folder strings to a follow-up

Keeps this change to the update-safe terminal host itself. The follow-up restores:
- the `stable-copy` spawner path class in daemon adoption telemetry;
- the per-terminal-host Full Disk Access row in Settings and its daemon RPC;
- writing folder usage descriptions into the app's Electron Helpers for the fork fallback.

The helper still takes its Desktop, Documents and Downloads descriptions from
`macos-folder-usage-descriptions.cjs`, so that constant stays.

* fix(i18n): restore en.json to this branch's base, not newer main

The previous commit took the file from a newer main, which pulled unrelated strings into this PR.

* refactor(macos): publish one terminal-host copy per helper build under a fixed launchd label

Name the copy after the helper's CDHash, reuse it when present, and
publish a new one by cloning into a staging dir, verifying it once
against the app's designated requirement and renaming it into place.
One launchd label per profile and protocol replaces per-launch labels,
so job.json, producer pids and per-label collection go away. Retirement
deletes any other copy lsof shows unused. The stable launcher returns
null for every fork fallback and logs its reason.

* fix(packaging): copy terminal-host dependencies from where Node resolves them

A nested package's own dependencies were never walked, so a hoisted
dependency of a nested package was left out of the helper. Resolve each
dependency through the node_modules lookup Node performs from its parent
and copy it to the same relative place, unless a copied package already
carries it.

* fix(packaging): fail release and CI mac builds that cannot boot a terminal-host slice

An arm64 runner without Rosetta skipped the x64 helper's PTY boot check
and shipped the slice untested. Release channels and CI now fail there;
local cross-arch builds still skip with a log line.

* docs(macos): describe the per-build terminal-host copy and its fallback reasons

* docs(macos): keep the codesign command on one line

* test(macos): type the launchd test state without an assertion

* docs(macos): name what the login-session flag gates in the stable launcher

* fix(macos): wait for launchd to drop a booted-out terminal job before reusing its label

---------

Co-authored-by: Neil <neil@stably.ai>
2026-10-11 14:54:36 -04:00

12 KiB
Raw Blame History

Design System

All UI work — layout, color, typography, spacing, component selection, UX behavior — must follow docs/STYLEGUIDE.md. Most of it is linted: pnpm run check:code-quality:changed fails on new restyles of a components/ui/ primitive, raw palette colors, and computed className strings; pnpm lint fails on any class Tailwind cannot generate. See the Enforcement section of the style guide before suppressing either. Use the tokens defined in src/renderer/src/assets/main.css (the canonical source) and the shadcn primitives in src/renderer/src/components/ui/. Don't invent new color values, font sizes, or shadow tiers when a documented one already covers the role. When STYLEGUIDE.md is silent, follow the resolution order in its final section.

Electron UI Validation

Always run tests and agent-launched apps in the background with ORCA_BACKGROUND_LAUNCH=1. Never steal monitor focus or reveal test windows: no show(), showInactive(), bringToFront(), app.focus(), or OS activation. Use CDP screenshots of hidden renderers. Keep native-focus and visible-window tests paused on the user's desktop; run them on an isolated display or CI. Rebuild modified launch-policy code before running an app; stale build wrappers are not safe.

Use the $electron skill and Playwright CDP for rendered Orca UI checks. Do not use computer-use for Orca UI validation.

Style

Reuse Before Reimplementing

Before writing new logic at any scale — a function, component, IPC channel, state store, or whole subsystem/flow — check whether an existing implementation already does the job (or nearly does). Extend or generalize it instead of building a parallel version; only write from scratch when nothing fits. Keep the check proportionate: a quick search for trivial code, a real one before building anything substantial.

Concise/Brief Non-obvious Comments ONLY

  • DO NOT: be verbose, explain the obvious, walk through the code ("WHY not HOW")
  • BE CONCISE. 1 LINE if possible

Lint Rules: Do Not Disable Max Lines

NEVER add a max-lines disable (eslint-disable max-lines, oxlint-disable max-lines, or line-specific variants), and never add a per-file max-lines bump in mobile/.oxlintrc.json.

File and Module Naming

Never use vague names like helpers, utils, common, misc, or shared-stuff for files, folders, or modules. They carry zero info and tend to become dumping grounds. Name files after what they actually contain — prefer the concrete domain concept (e.g. tab-group-state.ts, terminal-orphan-cleanup.ts) over the generic role (tabs-helpers.ts, terminal-utils.ts). If you find yourself reaching for helpers, the file probably has more than one responsibility and should be split, or there's a better name hiding in the code that describes what the functions operate on.

Type Declarations: Prefer .ts Over .d.ts

Type Assertions: Prefer Checked Types

Avoid type assertions except as const. Unavoidable casts need a line-specific SAFETY: explanation:

// oxlint-disable-next-line typescript/consistent-type-assertions -- SAFETY: Explain the verified invariant here.

Verifying Changes

  • Typecheck: pnpm tc (or tc:node / tc:cli / tc:web)
  • Test: pnpm test [path/to/file.test.ts]
  • Lint: oxlint, or pnpm run check:code-quality:changed for changed files (full pnpm lint is slow); format with pnpm format
  • Design system: pnpm run lint:design-system for the full renderer report (not a gate); the changed-lines gate above is what CI enforces
  • Real Claude CLI: when you change Claude structured-session code (src/main/claude/claude-structured-*), run ORCA_REAL_CLAUDE_CLI_TEST=1 pnpm test src/main/claude/claude-structured-real-cli.test.ts src/main/claude/claude-structured-real-cli-fold.test.ts; it uses your real Claude login

Writing Pull Requests

Fill in .github/pull_request_template.md, written for a reviewer who has never seen this code:

  • No jargon — plain language, no internal shorthand.
  • The before and after as the user experiences it.
  • The mechanism you changed, not just the symptom.
  • Why this approach over the alternatives you considered.

Cover all four concisely. Don't pad or walk the diff.

Considerations

Worktree Safety

Always use the primary working directory (the worktree) for all file reads and edits. Never follow absolute paths from subagent results that point to the main repo.

Cross-Platform Support

Orca targets macOS, Linux, and Windows. Keep all platform-dependent behavior behind runtime checks:

  • Keyboard shortcuts: Never hardcode e.metaKey. Use a platform check (navigator.userAgent.includes('Mac')) to pick metaKey on Mac and ctrlKey on Linux/Windows. Electron menu accelerators should use CmdOrCtrl.
  • Shortcut labels in UI: Display ⌘ / ⇧ on Mac and Ctrl+ / Shift+ on other platforms.
  • File paths: Use path.join or Electron/Node path utilities — never assume / or \.
  • Windows terminal shells: --shell picks the shell a terminal is; --command is typed into whatever shell the host spawned, so a shell choice routed through command silently becomes a child process.
  • Windows setup scripts: the setup/issue-command runner is a .cmd batch file unless the script starts with a #! line — never derive that from the user's terminal-shell preference, and never launch a .cmd runner with a bare cmd.exe /c from a Git Bash pane (MSYS rewrites the /c).
  • Windows child processes: start them through runProcess/spawnProcess from @orca/process-host — never child_process directly. It pins windowsHide, refuses shell: true, and encodes .cmd/.bat arguments so neither CommandLineToArgvW nor cmd.exe mangles them. Recognised npm/pnpm .cmd shims are resolved to their real target so the spawn skips cmd.exe entirely.
  • Ripgrep: Orca bundles rg for every platform, WSL, and SSH remotes. Spawn it through spawnBundledRipgrep (main) or resolveRelayRipgrepCommand (relay), never a bare 'rg' — Windows resolves a bare name in the spawn cwd before PATH. Don't add git/readdir fallbacks locally; the relay's chain exists only for hosts an upload never reached.
  • Windows process enumeration: read the table through src/main/windows/windows-process-table.ts, never by forking powershell.exe.
  • Windows MSYS/Git Bash panes: their children break away from the per-PTY job unless it is created without JOB_OBJECT_LIMIT_BREAKAWAY_OK, and a conpty.node built before that fix passes every existing gate.
  • Windows daemon-host relocation: the terminal daemon runs from a copy of the app runtime under %LOCALAPPDATA%, which is what survives an auto-update.
  • macOS daemon launch: packaged macOS runs the terminal daemon from a copy of the signed terminal host helper through launchd, so an update cannot sever its privacy grants. Before touching that copy, its cleanup, or the fork fallback, read docs/reference/macos-daemon-tcc-attribution.md.
  • Windows EDR signal: don't add -ExecutionPolicy Bypass, -EncodedCommand, cmd.exe /c with escaped free text, per-operation interpreter spawning, or runtime Add-Type compilation without reviewing the EDR impact first — behavioural EDR scores each of those, and being signed does not clear them.
  • WSL commands: build argv with buildWslExecArgs (always --exec — under --, wsl.exe expands $name in every argument and silently rewrites the script), and fence anything whose stdout you parse with buildWslCapturedLoginShellCommand, because the interactive login shell prints the distro banner to stdout.
  • Linux native modules: keep the glibc floor at Ubuntu 20.04 / glibc 2.31. A module compiled from source on a newer runner can reference symbol versions absent on the floor and crash the app on startup. Packaging fails if a bundled native binary needs newer glibc.

Native Dependency Installs

Ordinary pnpm install covers the host OS and CPU only. Before packaging for another architecture — including pnpm build:mac, which builds x64 and arm64 by default — run pnpm install:release. electron-builder only warns on a missing extraResources source, so the beforePack guard is what turns a thin install into a build failure instead of a silently broken artifact.

SSH Use Case

All changes must consider the SSH use case. Don't assume local-only execution. Before changing anything that reports on, stops, or lists remote work, the execution host owns everything that touches execution, and loss of contact is never evidence of process death — the verdict vocabulary is live / unverifiable / exited, with no synonyms.

Folder Workspace Use Case

All changes must consider folder workspaces as well as git worktrees. Don't assume every workspace is a git worktree.

Agent Status

The execution host owns agent status in one store, the hook server's, and every reader (sidebar, worktree ps, mobile, dashboard) subscribes to it. New producers write into that store, and readers keep only presentation policy.

Agent Terminal Screens

A rule that reads what an agent CLI paints on a terminal — readiness, blocked prompts, idle — must be written against a captured transcript, not a remembered screen. Use config/scripts/capture-agent-pty-transcript.mjs to keep escapes and wrapping intact and scrub account identifiers before they reach git.

Remote Wire Compatibility

Clients and remote Orca servers update independently, so mixed versions are the normal state. Before changing anything a paired client and host exchange — RPC params, stream frames, or the content either side publishes over them — preserve compatibility with older peers. A new optional field is safe; a new stream opcode must be capability-negotiated because decoders drop unknown opcodes silently; and changing what the host publishes reaches old clients even with no wire change.

Git Binary Compatibility

Orca runs the user's Git binary on native, WSL, and SSH hosts, which may all have different versions. Treat Git 2.25 as the core-workflow baseline.

When adding or changing a Git command:

  • Check when every subcommand and option was introduced. For newer behavior, keep a baseline-compatible fallback or degrade safely.
  • Use GitCapabilityCache with a narrow unsupported-error predicate so recurring operations do not retry a known-invalid command. Do not rely only on git --version; wrappers such as simple-git do not remove host-version differences.
  • Scope capability state to the host that executes Git: native, WSL distro, SSH provider, or relay connection. Cover the first fallback, later cached calls, concurrent probes, and relevant host isolation in tests.
  • Keep the real-binary compatibility contract in PR CI current. When adopting a newer Git feature, add its version boundary so the preferred command and fallback both run against representative Git releases.
  • Preserve commands that begin with global Git options such as -c before the subcommand, including auto-maintenance suppression used by worktree-create fetches.

Git Scan Safety

  • Never enumerate every ref and then run git ls-tree -r or git show once per ref. That ref × tree fan-out can retain gigabytes of output before a downstream sort -u or search can make progress.
  • Prefer rg over the checked-out files for source searches. For history or refs, use a named ref, an explicit namespace/path, --max-count, and a bounded output; do not use an unqualified --all scan as a first diagnostic.
  • Keep repository-wide commands targeted to the current repository and worktree. If an unbounded scan is genuinely required, measure the ref count first, explain the cost, and get confirmation before running it.

Git Provider Compatibility

Source-control and review changes must consider GitLab and other supported git providers, not only GitHub. Keep provider-specific behavior behind explicit checks, and avoid GitHub-only naming for generic review concepts.

GitHub CLI Usage

Be mindful of the user's gh CLI API rate limit — batch requests where possible and avoid unnecessary calls. All code, commands, and scripts must be compatible with macOS, Linux, and Windows.