Commit Graph
49 Commits
Author SHA1 Message Date
Neil 815854777b refactor: give browser, RPC and process contracts concrete owners (#40167)
* refactor: separate execution route and child exit owners

* refactor: move browser data into defining feature owners

* refactor: give runtime RPC entry its shutdown implementation

* refactor: use defining process types and colocate host telemetry test

* docs(process-host): point piped child contract to Node owner

* docs(process-host): clarify both piped contract owners
2026-11-08 23:32:21 -08:00
Neil e9f4df2c0f refactor: organize rate-limit providers and remove forwarding modules (#40116)
* refactor: group rate-limit providers under concrete owners

* test: preserve mixed Claude account-switch regression

* refactor: consolidate Antigravity service contract import
2026-11-08 20:29:13 -08:00
Neil 51fcb15154 refactor(browser): organize automation, cookies and session policy owners (#40002)
* refactor(browser): give agent-browser implementations concrete owners

* refactor: give cookie import and storage policies concrete owners

* test: target the concrete Chromium registry edge in cookie fixtures

* refactor: organize browser profile identity and permission owners

* refactor(browser): import input and cookie types from defining owners
2026-11-08 16:04:41 -08:00
Neil b5e38ad9c6 refactor: organize agent hook installation and server owners (#39969)
* refactor: organize agent hook installation and server owners

* refactor: align hook mock registries and packaged entries
2026-11-08 13:23:03 -08:00
Neil 0d1220c31f Organize SSH and managed-server modules by concrete owner (#39937)
* refactor(main): organize SSH transport and managed server owners

* Remove orphan deploy readiness fixture and shrink test baselines

* refactor(wsl): import lock records from their defining owner

* refactor(wsl): use the canonical deferred update contract
2026-11-08 12:20:22 -08:00
Neil 3a0e9a27d0 refactor: organize backend runtime by defining owner (#39625)
* refactor(main): organize runtime implementation owners

* test: retire wall-clock CLI title recovery case

* refactor: route new folder mutation to runtime store owner

* test: import folder mutation store contract from its owner
2026-11-07 22:24:34 -08:00
Neil 2f217b4d0d refactor(daemon): group endpoint, provider, session and transport owners (#38462)
* refactor(daemon): group transport, provider, endpoint, and terminal owners

* refactor(daemon): consolidate owner imports and preserve pinned baseline sources

* refactor(daemon): import termination budgets from their defining owner

* fix(ci): retain serve mode coverage for extracted orcad startup

* fix daemon build smoke to load concrete argument owner

* read daemon protocol facts from historical defining paths

* route protocol revision owners through compatibility checks
2026-11-05 01:54:51 -08:00
Neil 7bf0377072 refactor: place terminal and workspace contracts with concrete owners (#38061) 2026-11-03 22:49:34 -08:00
Neil ec3a707868 refactor(browser): group routing and client-host owners (#37872) 2026-11-03 12:28:16 -08:00
Alonso AlmeidaandAlonso Almeida 3b91fe4971 fix(cli): forward OpenCode --method and --answer from account add (#37622)
Forward OpenCode login methods and form answers into private account enrollment. Reject missing repeated values before login and preserve literal arguments through the common Windows process launcher, with terminal input, JSON output and cancellation retained.

Thanks to Alonso Almeida (@Alonsojrr14) for the original implementation and follow-up fixes. Tracked in #37621.

Co-authored-by: Alonso Almeida <alonsojr1269@gmail.com>
2026-11-03 01:12:19 -08:00
Neil 5d5b4f4f3d refactor(ai-vault): organize provider scanners and history owners (#37742) 2026-11-03 01:05:52 -08:00
Neil f3088d53c9 refactor(main): group terminal support by lifecycle owner (#37660) 2026-11-02 22:13:39 -08:00
Neil 7983d5064c refactor: group agent launch and workspace hook owners (#37612)
* refactor: group agent launch and workspace hook owners

* refactor: align trust CI scope and reliability paths
2026-11-02 17:51:53 -08:00
Neil b89cfbd09d refactor: group desktop integration and lifecycle owners (#37600)
* refactor: group desktop integration and lifecycle owners

* refactor: preserve runtime literals and mock owner boundaries

* refactor: link historical reliability evidence to current test paths

* refactor: retain the historical lid-sleep baseline path
2026-11-02 17:08:59 -08:00
Neil c550fbca06 refactor(browser): group agent browser adapter modules by owner (#37223) 2026-11-01 15:32:34 -08:00
Neil 520588fe97 refactor(codex): group app-server transport and lifecycle owners (#37109) 2026-11-01 11:05:38 -08:00
Neil 756d623955 perf(terminal): bound the in-process macOS process table's blocking calls and ps fallbacks (#31783) (#36282)
* fix(terminal): bound the in-process macOS process table's blocking calls and ps fallbacks (#31783)

Follow-up to #35987 from its post-merge reviews.

- cwd and executable path: node-pty's readProcessCwd/readProcessPath now run
  PROC_PIDVNODEPATHINFO and proc_pidpath on a libuv worker (the synchronous
  exports are gone, so an older pty.node falls back to ps/lsof). JS caps them at
  two outstanding lookups and 250ms; a wedged mount or a full cap answers from
  lsof -b, which never stats, so a dead NFS/SMB/FUSE mount cannot freeze the
  daemon's event loop or take the whole pool.
- Unreadable argv prints "(comm)", as ps does, instead of a proc_pidpath lookup
  under Orca's credentials on the table worker.
- The table exports each row's euid. Argv recovery runs setuid ps only for
  another user's processes, never for our own mid-exec or exiting ones (or
  SIDL/exiting rows), which ps cannot read either: a pane looping short
  commands forked ps on ~15% of full reads, now none.
- Recovery is coalesced per process across concurrent readers and a failed
  recovery waits 5s before costing another ps. The cache keys on comm and start
  time.
- Full-table and shell-foreground captures, which feed strict evidence with
  the capture's age, take only argv ps read after their request; display
  readers keep the 30s cache.
- While the reader is down (stalled-pool breaker open, or a rejecting addon,
  which now backs off 1s..30s), runPsInProcess runs /bin/ps itself through one
  shared budget: identical invocations share a run, a burst of 4 forks then one
  a second, and a run spawned before a request never answers it. Descendant
  teardown and foreground-group signalling keep their own full-timeout ps
  (callerRunsPs). 13 pollers at 750ms: 1,040 ps/min before, 63 after.

No daemon protocol bump: the wire format is unchanged, and each daemon loads
JS and pty.node from the same runtime copy.

* fix(terminal): make the libproc reader survive worker teardown, fail closed and share one fair fork budget (#31783)

Review follow-ups to the previous commit.

- Native: every libproc AsyncWorker derives from OrcaPromiseWorker, which skips
  JS completion once an env cleanup hook marks the env closing and swallows a
  Napi::Error thrown mid-teardown. Terminating a worker_thread with cwd or table
  reads in flight aborted the process; 200 terminates each now survive.
- (comm) in a terminal's foreground group (argv neither reader recovered, e.g.
  a root `su`) judges 'unknown', never 'agent', so the launch line is not typed
  into it.
- Fresh argv only for getFreshSnapshot captures; the 500ms TTL readers keep the
  30s cache. Argv recovery runs through its own fork budget (2, then one per
  2s). Root `su` pane, 13 pollers at 750ms: TTL 79 -> 2 ps/min, fresh 158 -> 31.
- The fork budget serves invocations in FIFO turns, grouping each one's waiters
  onto one run. 13 streams: served 80,42,1,1,2,0x8 -> 10-12 each.
- Forced subprocess captures (a held naming pass) go through the fallback
  budget: 5 readers, 400 direct ps/min -> 63, 395 of 400 captures served.
- A failed recovery is retried 5s after ps gave up, not after it started.
- Tests pin Orca's euid instead of reading the runner's and assuming root is 0.
- process-cwd keeps one lsof per pid until it exits, so a dead mount no longer
  leaks one per poll; shell readiness reads the exe path through runProcess,
  which gives up on an lsof that ignores SIGTERM.
- A #35987 pty.node (table reader, no async path reads) keeps the in-process
  table; cwd/path use lsof, with a one-time warning.
- captureProcessTable moves to process-table-ps-capture.ts.

* fix(terminal): make withheld argv unverifiable and reuse recovered argv for at most 2s (#31783)

Round-2 review follow-ups.

- Worker teardown: drop OrcaPromiseWorker and its env flag (Node drains the
  pool before cleanup hooks, so it never fired). node-pty's mac pty target now
  builds with NODE_API_SWALLOW_UNTHROWABLE_EXCEPTIONS, which covers both the
  call-in and completion paths. The smoke test now calls the reads from a
  timer while terminating: the previous build aborted 2 of 3 runs, this one
  survives 3/3 x 100. pty.cc is back to the first follow-up's exact content.
- A foreground process whose argv ps withheld (`(comm)`: another user's,
  exiting, or a recovery the budget skipped) makes the batch resolver, and so
  host-stamped evidence, `argv_unreadable`/unverifiable rather than live with no
  agent; agent-foreground resolution reports it unavailable.
- judgeTerminalForeground fails closed only for an unreadable shell (`(bash)`
  from a root `su`), not our own exiting `(git)`.
- Recovered argv is reused for at most 2s per identity (pid, start, comm) for
  every reader (the next commit ages evidence by it), and
  fresh reads no longer queue behind the recovery budget. Root `su` pane, 13
  fresh pollers: p50 2,989ms -> 53ms, 31 -> 26 ps/min (TTL reader 1 -> 26).
- The fork budgets serve interactive and fresh readers before display polls
  (cheap and TTL snapshot captures), FIFO within each. 20 display streams plus
  a 3s launch-paste read every 5s: 0/12 -> 12/12 served, polls 4-7 each.
- On a selected terminal, the foreground group's argv is recovered even with no
  process of ours left (`exec su`); the root login is not.
- Budgeted captures keep the 32 MiB output ceiling.

* fix(terminal): age strict evidence by its oldest recovered argv (#31783)

Round-3 review follow-ups.

- Recovered argv carried no read time, so a capture reusing it (cached, or a
  recovery still in flight) stamped its own capture time over argv up to 2s
  older, and the remote write guard could take a same-comm re-exec for the
  agent. Rows now carry when ps read their recovered argv; runPsInProcess
  reports the oldest through onArgvReadAt, and the snapshot reader stamps
  capturedAtMs at that input, so capturedAgeMs is the age of the oldest input
  and consumers' own freshness checks (the write guard's 500ms) decide. A
  recovery begun more than 2s before a reader asked no longer answers it, and
  an older recovery finishing late cannot replace a newer one.
- resolveSpawnFileForegroundFromRows reports a withheld `(node)` foreground
  unavailable.
- One budget queue entry per invocation: an interactive request promotes a
  queued poll's, a poll joins an interactive one.
- parseShellForegroundRows gets its JSDoc back.

* fix(terminal): keep a cached name from authorizing a refused write; integer, under-2s argv ages (#31783)

Round-4 review follow-ups.

- When the relay's evidence declines a write (stale, unverifiable, another
  agent), its foreground name may still prove the shell but no longer proves
  the agent: that name can be the tracker's 1s display cache, e.g. `claude`
  cached just before Claude exited to a root shell under su. Relays that send
  no evidence keep the name as before.
- Snapshot ages round up to whole milliseconds; the wire rejects fractions
  from the monotonic-to-wall conversion.
- Recovered argv is reused for at most 1.5s, so reuse plus the 500ms snapshot
  cache stays under the 2s freshness gate; reuse ages from when that ps
  spawned, not from the wait for its fork token, and only an answer is
  remembered (a timed-out or withheld run proves nothing). Fork budget
  unchanged. Root su pane, 13 pollers: 30 ps/min; fresh p50 51ms; strict
  evidence age p50 752ms, max 1,508ms.

* fix(terminal): keep the relay's agent name behind structurally unverifiable evidence (#31783)

The previous commit refused the name fallback whenever the relay sent any
evidence, so `unverifiable` evidence for structural reasons also blocked
writes: npm Codex (node codex.js spawning native codex) and Gemini (relaunching
itself) put two agent rows in the foreground group (ambiguous_foreground_group),
and remote tmux reads multiplexer_boundary. On SSH POSIX hosts that dropped
launch prompts and failed worker briefs, conversation send/cancel and
automations with agent_not_in_foreground.

The name now cannot authorize only when the evidence saw the group (`live`:
stale, or no agent in it, as in the cached-`claude`-over-root-shell repro) or
withheld a foreground argv (`argv_unreadable`, possibly a root `su` shell).
Other unverifiable reasons keep the name fallback, as before this branch.

* fix(terminal): let a relay name complete only fresh agent-positive evidence (#31783)

The SSH name fallback still let a stale or cached name authorize a write under
any unverifiable reason but argv_unreadable: ps EAGAIN (process_table_unreadable)
after the tracker cached `claude` over a root bash, anchor_missing,
tty_boundary (`sudo -i` with use_pty, a root shell on a new pty), or a 1.5s-old
ambiguous capture.

- The name now completes evidence only when it is fresh (within the guard's
  500ms bound), its reason is one that saw an agent row
  (ambiguous_foreground_group, multiplexer_boundary,
  candidate_start_time_missing), and the name is the launched agent. Every other
  reason, including ones this build does not know, reads unknown. Relays that
  send no evidence keep the name, for mixed versions.
- The relay's evidence names one agent that holds two rows of the group (npm
  Codex: node codex.js spawning native codex; Gemini relaunching itself), so a
  new relay proves them live even when its name is `node`. Two different agents
  stay ambiguous. No wire change.

* fix(terminal): judge a relay name by the host's capture age; keep old relays' npm Codex (#31783)

- Name completion compared the capture's start with the read's start, so a
  slow response made old evidence look fresh: a capture 1,500ms old arriving
  after 1,100ms passed. It now uses the host-reported capturedAgeMs against
  the 500ms bound, which latency cannot shrink. Live evidence keeps its
  capture-start rule.
- Relays before same-agent dedupe report npm Codex (node wrapper plus native
  binary) as ambiguous_foreground_group and name it `node`. For that reason
  only, `node` now completes the evidence when the launched agent is
  npm-installed per the node entrypoint tables; other agents and reasons stay
  fail-closed.

* fix(terminal): bound relay-name completion by twice the snapshot TTL (#31783)

The 500ms bound refused ~9.5% of tmux name completions on a host with 13
pollers (cached snapshot up to one 500ms TTL old plus ~30ms capture, max
528ms), and a refusal drops the launch prompt or throws with no retry. The
bound is now 2 x PROCESS_TABLE_SNAPSHOT_MAX_STALENESS_MS (1s), still on the
host-reported age: 0 of 767 reads exceed it, the delayed 1,500ms repro still
refuses, and the cached-name hazard (the tracker's ~1s display cache) is no
wider.

* fix(terminal): complete relay evidence with a name read now, not the display cache (#31783)

With the 1s bound, a capture at 0-528ms, `codex` cached at 100ms, Codex
exiting at 101ms and the guard at 999ms still let the cached name authorize a
write over the bash that replaced it.

- Name completion (structural reasons, host age within 2x the snapshot TTL)
  now asks confirmForegroundProcess, which reads the foreground now (a fresh
  scan, or node-pty's raw process name) rather than the tracker's 1s display
  cache; a null answer reads unknown. No new wire surface: the local daemon answers
  confirmForegroundProcess, paired runtimes evaluate guards locally, and the
  production controller returns null (read as unknown, failing safe) when a
  provider lacks it. Controllers without the method, and relays sending no
  evidence, keep the cached name as the legacy path.
- Live evidence also accepts a host-reported age within 2x the TTL, so a real
  agent behind a 500-528ms cached snapshot on a fast link is no longer
  refused; a capture begun after the read was asked still counts, and
  5,000ms-old evidence still refuses.

* fix(terminal): confirm live evidence older than one snapshot TTL with a fresh read (#31783)

Accepting live evidence up to 2x the TTL reopened the window the previous
commit closed for name completion: a capture at 0-528ms, the agent exiting at
~529ms, and at 999ms `live claude` still read as agent over a bash.

Live evidence begun no more than one TTL before the read was asked (so any
host age within one TTL, however long ps took) still answers alone. Between
that and 2x the TTL, confirmForegroundProcess reads the foreground now: the
launched agent proves it, a shell reads shell, anything else unknown.
Controllers without that read fall through and refuse, as before.

Reachability: the remote (host.remote) write-guard path has no production
caller since the SSH provider's removal (#27499); the registry dispatches only
local providers and paired runtimes evaluate guards locally. This is defence
in depth for any future remote provider.
2026-10-30 01:05:34 -07:00
OrcaWinandOrcaWin 2d6ee2ab9b fix(emulator): run Android commands through the process host (#35599)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-10-29 15:00:48 -07:00
Neil 525719789f perf(terminal): read the macOS process table in process instead of forking setuid ps (#35987)
* perf(terminal): read the macOS process table in process instead of forking setuid ps (#31783)

/bin/ps is setuid root, so every exec costs an uncached syspolicyd
assessment and an EDR inspection. The terminal host forked ~7-10 ps/lsof
per second with 4 agent panes, and each stale daemon multiplied that.

node-pty's patched pty.node now exports processTable (sysctl
KERN_PROC_ALL/KERN_PROCARGS2 on a worker thread), processCwd
(PROC_PIDVNODEPATHINFO) and processPath (proc_pidpath). runPsInProcess
renders the same ps column text for the invocations Orca issues, so the
existing parsers and callers are unchanged; anything it cannot reproduce,
non-macOS hosts, and builds without the exports keep running ps/lsof.

Measured with 4 anchored panes polled every 750 ms for 60 s: ~440
ps/lsof children before, 0 after.

* fix(daemon): roll the in-process macOS process table into a fresh v54 daemon (#31783)

* fix(terminal): bound, escape and complete the in-process macOS process table like ps (#31783)

Review follow-ups for the libproc reader:

- Timeouts: every caller passes its ps timeout; a read stalled behind a
  saturated libuv pool resolves timed out (code null, timedOut) at that
  budget, exactly as the killed ps would. Identical native reads share one
  worker, each caller racing its own deadline; past 8 distinct reads in
  flight callers run ps. execFile-shaped callers reject like a failed child.
- tty naming shares one slot with captureDarwinProcessTable's guard, so a
  draining naming pass sends both paths to ps instead of naming again.
- Commands: the addon renders argv with adv_cmds' getproclline and strnvis
  (VIS_TAB|VIS_NL|VIS_NOSLASH) under the caller's LC_CTYPE via a
  thread-scoped uselocale, so newlines print as \012 instead of splitting
  tokens; zombies print <defunct>.
- Other users' processes: unreadable argv is marked commandIncomplete. Only
  -p selections and processes on a terminal under the user's own readable
  process there (sudo in a pane) get a targeted `ps -p ... -o pid=,command=`,
  cached by (locale, pid, start time); root daemons and a pane's root login
  never spawn ps.
- -t selects with KERN_PROC_TTY after ps's addelem_tty path resolution, so
  per-terminal reads skip argv for every other process (~0.3 ms vs 4 ms ps).
- -p deduplicates; rows sort by terminal then pid as pscomp(); stat adds
  ps's < N X V L modifiers. The export is renamed readProcessTable so an
  older pty.node falls back to ps rather than misreading the new query.

* test(terminal): diff the real libproc addon against /bin/ps on macOS (#31783)

Spawns a pty holder whose argv carries a newline, tab, invalid UTF-8 and é,
then compares Orca's ps invocations column by column: pid/ppid/pgid/tpgid/
lstart/tdev/tty exactly, commands for the user's own processes, and stat
modulo running/idle (thread state needs setuid ps). Rows are compared only
when two in-process reads around the ps run agree, so churn cannot flake it.

* fix(terminal): never serve a native process table read begun before the request (#31783)

A timed-out read stayed shareable, so a later descendant capture could join
an earlier poll's abandoned worker and take its rows as fresh; a descendant
that exited meanwhile could have its reused pid SIGTERMed. The addon now
reports when its worker actually began reading (startedAtMs), and
runPsInProcess refuses a shared read that began before its own request and
reads again, so every consumer that signals from rows (descendant
teardown, foreground-group signalling, identity probes) only sees rows
scanned after its stamp. Regression: pty-descendant-termination-libproc.

Also from review:
- argv recovered with setuid ps is keyed by pid, start time to the
  microsecond, comm and executable path (exec changes the latter two) with
  a 10 min TTL and 512-entry bound.
- Recovery covers a sudo use_pty tree: the unreadable chain must start on
  the terminal of the user's own readable process, so the root shell and
  agent on sudo's own pty qualify while root daemons, GUI helpers and a
  pane's root login (own pty, under a terminal-less or tty-inheriting host)
  do not.
- Concurrent tty= readers wait out the single naming pass, and readers past
  the 8-read cap queue behind one, within their deadlines, instead of
  bursting /bin/ps at startup restore (24 panes polled at once: 0 ps).
- darwin-terminal-naming-slot.ts is now darwin-terminal-naming-single-flight.ts.

* test(terminal): require /bin/ps row order in the libproc parity test (#31783)

Parsers that take the first matching row depend on ps's terminal-then-pid
order, so compare the order of the rows both captures share, not just their
contents.

* test(terminal): pin libproc read start times instead of reading the wall clock (#31783)

* fix(terminal): match native process table reads to requests by sequence, not wall clock (#31783)

Freshness compared the worker's gettimeofday with Date.now(), so a clock
step back let an abandoned read's rows through (wrong-pid SIGTERM) and a
step forward could reject fresh reads until the deadline. Each read now
takes a JS-side sequence number when created and is submitted a microtask
later; a caller joins only a read not yet submitted or created after it
registered, so callers in one turn share a worker and nothing submitted
before a request serves it. The addon drops startedAtMs again. Deadlines,
timeouts and the argv cache use performance.now().

Callers past the 8-read cap wait on one slot notifier and each joined read
has one settle notifier; both unsubscribe on timeout, so a stalled pool
retains no per-caller reactions. If the pool is still at the cap when an
overflow caller's deadline passes, the reader is marked stalled for 30s
and callers run ps, as before this branch.

Recovered argv is trusted for 30s: exec can keep pid, start time, comm and
executable (node -> node with another script) and no cheap field shows it,
so a sudo pane costs one targeted ps per process per 30s.

* fix(terminal): keep ps fallbacks off a stuck terminal naming pass (#31783)

- captureDarwinProcessTable's canonical fallback, taken while the naming
  slot is busy, now runs /bin/ps directly instead of re-entering the
  in-process reader and waiting on that same slot (a held lstat timed out
  every strict capture without one subprocess fallback).
- In-process tty= readers wait out a short pass (startup restore) but hand
  a pass older than 500ms to ps, which names terminals out of process.
- The naming single flight keeps one settle notifier per pass; waiters sit
  in a Set and drop off on timeout, so timed-out callers leave no reactions
  on a stuck pass (was one per call). The notifier moves to
  deadline-settle-notifier.ts, shared with the native read scheduler.
- Once the native read breaker has tripped, a caller finding the cap still
  held after the cool-down returns to ps at once instead of waiting out its
  timeout.
- Descendant teardown and foreground-group signalling give the native read
  half their budget and fall back to a fresh ps on a native timeout, so a
  partly stalled pool cannot cost a kill.

* fix(terminal): give signalling ps fallbacks their full baseline timeout (#31783)

Splitting the budget in half shrank ps's window behind a stalled native
read: a 750ms ps lost descendant SIGTERMs and an 80ms foreground ps lost
group SIGWINCH. Descendant teardown and foreground-group signalling now try
the native reader for min(150ms, timeout/4) (inProcessSliceMs), then give
ps the whole original timeout; readProcessTableBeforeDeadline extends the
production reader's outer deadline by that slice, so ps succeeds exactly
when it did on main. The stale-read regression test now also asserts the
fresh rows' real descendant is signalled.

* fix(terminal): carry the in-process deadline allowance through every descendant reader (#31783)

readProcessTableBeforeDeadline granted the native slice only when the
reader was readProcessTable by identity, so the daemon shutdown sweep's
sharing wrapper kept a bare 250ms deadline: a stalled native read plus a
200ms ps missed it on every retry and descendants survived shutdown.
ProcessTableReader now declares deadlineAllowanceMs; the fresh reader sets
it (0 without the addon or with a stalled pool, so Linux and ps-only hosts
keep their exact deadlines), createProcessTableSnapshotReader and the
shutdown wrapper inherit it, and injected readers default to none. The
fresh reader moves to pty-descendant-fresh-table.ts. The real-ps parity
test gets a 30s per-case timeout, since whole-table /bin/ps runs slow
under suite load.

* chore(terminal): move the descendant reader's child_process baseline entry with it (#31783)

* test(terminal): pin the libproc test locale so lstart doesn't follow the runner's LANG (#31783)
2026-10-29 09:51:53 -07:00
Neil e0f05c3fc9 fix(wsl): Codex WSL hook installs and host hook checks never start a stopped distro (E9) (#35827)
* fix(wsl): Codex WSL hook installs and host hook checks never start a stopped distro (E9)

(a) installManagedHooksIntoWslRuntime checks the settled running gate inside
its trust queue right before the first UNC read, and again before the
post-hash fallback trust write; a later reconcile checks before revoking or
reinstalling. A stopped or unverifiable distro skips (null, as with no WSL
plan). The canonical-path probe runs through process-host behind
blockedWslDistroForExec; a refused probe settles as unavailable and keeps the
known identity. The guest shell preflight and a pane launch's home preparation
run as user acts on their distro.

(b) repo.hooksCheck takes an optional background flag (older hosts ignore it);
the renderer sends it, and the host refuses a stopped distro's background
check the same way the desktop's hooks:check does.

* fix(wsl): a pane launch is not a user act for the Codex hook install

A passive restore or remount reaches launch preparation too (#35813), so it no
longer runs as a user act; a stopped distro skips the install and the relay
installs once the distro runs. A skipped WSL home now returns not_installed
instead of falling through to the host writer, which would have written the
WSL home over UNC.

* fix(wsl): a fallback grant skipped mid-install reports trust pending

hooks.json is written before the hash wait, so a skip after it is not a clean
skip: return an error status saying the trust is pending instead of null.

* chore: drop the stale process-host direct-import baseline entry
2026-10-29 02:52:01 -07:00
32e804dd6f fix(wsl): every local Git call follows its project's runtime unless it says why not (D1c-6) (#33531)
* fix(wsl): every local Git call follows its project's runtime unless it says why not (D1c-6)

* fix(wsl): rebuild Git routes on the store's own writes, never on a timer, and route a new worktree by its link

* fix(wsl): route WSL-made worktree links, aliased checkouts and sync Git; host Git only where marked; no wsl.exe storms

* fix(wsl): map WSL-made worktree links through each project distro's automount root

* fix(wsl): arm the missing-distro cache only with the Git router on, and feed it spawn failures

* fix(wsl): keep synchronous Git host-only, read the setup runner's path from the link, and keep the sync launch under the ratchet

* chore(wsl): move the sync Git runner's process-host import allowance with it

* fix(wsl): read only per-worktree git paths from the link, and only from a git dir that still exists

* fix(wsl): name which Git ran on stream and spawn spans too, so status shows its route

* test(wsl): setup runner lookups declare host Git; cover the link read with the router on and off

The hooks tests expected the setup runner's `rev-parse --git-path` call to
carry `wslDistro: 'Ubuntu'` for a WSL-runtime project. On main that field
never reached Git: main's gitExecFileSync options have no wslDistro and it
calls resolveCommand('git', args, cwd) with no distro override, so the
lookup ran host Git (WSL only when the cwd itself is a WSL path). The tests
pinned call arguments, not what ran. The sync runner now states that with
`hostGit: true` and resolves exactly as main did, so the tests assert that.

New worktree-runner-script-link tests cover the router-gated link read:
router on reads the link; router on with no usable link falls back to host
Git and converts its Linux path; router off makes main's host Git call and
never reads the link.

* test(git): read src/main once in the host-Git routing guard, so it holds under load

* test(wsl): pin the setup runner's sync Git spawn to main's exact launch on POSIX

The setup runner's path lookup now passes hostGit: true where it passed a
wslDistro that gitExecFileSync never read, so the binary, args, cwd and env
it spawns are unchanged. Update the call-shape assertions and pin the POSIX
execFileSync call itself.

(cherry picked from commit 6aae6d4d090e13b75c1d9ad2734f2f717dde7570)

* fix(wsl): await WSL Git admission in spawns only when there is one

With the Git router off admitWslGit returns null; awaiting it anyway added a
microtask to every spawn's admission path that main doesn't have.

(cherry picked from commit 883aaeb90bc23e89cafb039708a565ab3d72620e)

* fix(wsl): keep main's Git scheduling when the D1c Git router is off

The remote-identity sweep moved every user's probes to the background
admission tier, which only the router reads; it now keeps its tier unless the
router is on. Admission wrapping and the spawn stderr tail also step aside
without the router, so those calls run exactly as on main.

(cherry picked from commit 151a3091307c03078d0c0b98487bbe3bf467ec27)

* test(wsl): point the distro allowlist test at the sync runner's new home, as host Git

Co-Authored-By: Claude <noreply@anthropic.com>
(cherry picked from commit 255e85fc93deba83addc4370f889c1b721180188)

* test(wsl): installing the Git router with the gate closed or off Windows leaves none installed

Co-Authored-By: Claude <noreply@anthropic.com>

* perf(wsl): skip the span-only Git resolution on every spawn while the Git router is off

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(wsl): read a distro's automount root only from a running list probed after its stop settled, so it never boots a terminated distro

---------

Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
2026-10-29 01:52:18 -07:00
edc0373fcd plugins: Plugin Alpha (#35470)
* i18n(plugins): drop retired plugin strings from the Portuguese catalog

* i18n: keep the empty project group in the Spanish and French catalogs

* i18n: retire translated entries for plugin strings this change rewords or removes

* plugins: syntax contributions

A plugin can contribute up to 16 syntaxes: a TextMate grammar, optional
language configuration, and the file extensions and names it claims.
Main validates every grammar when the plugin is approved (shape, scope,
size, nesting) and lists only the file associations. A language's
grammars cross IPC only when a matching file opens, and includes must
resolve inside the same plugin. Catalog releases can declare syntax.

* plugins: highlight files with plugin syntaxes

Files Orca shows as plain text use a matching plugin syntax. Its grammar
compiles once in a worker that stays warm between edits and runs under
core's per-line limits; a worker that stops reporting progress or crashes
is terminated, the language stays plain, and the plugin card says so.
Tokens reach Monaco as semantic tokens, so Monaco asks only for models
shown in an editor, keeps one request per model, and keeps painting the
previous colors, shifted by edits, until the new pass lands. Plugin
languages borrow ids from a pool of fixed Monaco languages, counted by
the editors using them; a language no editor uses moves its models back
to plain text, and the least recently used idle slots are reused first.

* plugins: Gleam syntax example pack

An official-style data plugin with the upstream Gleam TextMate grammar
(Apache-2.0, provenance and license alongside) and language
configuration. A test installs it through the production inspection and
loads it through the syntax registry.

* plugins: list themes, icon themes and syntaxes in the marketplace preview

The preview's Includes line named only language packs, commands,
shortcuts, VM recipes, panels and events, so a theme or syntax pack read
as "Manifest metadata only".

* fix(plugins): a syntax failure belongs to one revision, not the syntax id

A grammar that failed in revision A kept the plugin card flagged after revision B
replaced it. Failures are now keyed by id@revision, matching the runtime slots,
and dropped when the installed list no longer contains that revision.

* chore(plugins): format the Gleam example grammar

* plugins: import the Monaco editor API, not the full bundle, from the syntax runtime

* plugins: type the syntax engine against the Monaco API it uses

* plugins: suggest a catalog plugin for files Orca can't open

When a file shows as plain text or as a binary Orca can't display, and a
cached catalog lists a release whose fileTypes include its extension, the
editor offers it: View plugin opens the marketplace preview (never an
install), Not now hides it for the session, and Don't suggest remembers
the extension. Suggestions come only from cached, verified catalogs, skip
kill-listed releases and anything an installed (even disabled) plugin
already handles, and can be turned off in Plugins settings. A release's
fileTypes must match its manifest's syntax extensions.

* fix(plugins): scope the file-type suggestion's Not now to the file type

The editor stays mounted when its file changes, so a single hidden flag
suppressed the suggestion for the next file type too. Track which file
type was hidden instead.

* plugins: type the plugin-owned settings normalizer inputs

* plugins: update the preview-request callback ref after render, not during it

* test(plugins): smoke specs for themes, syntaxes and file-type suggestions

Three end-to-end specs join the demo and marketplace journeys:
- an app theme and icon theme apply from Settings and fall back when the
  plugin is disabled;
- a plugin syntax highlights an already-open file, re-highlights it after a
  grammar update, and lets go when disabled;
- a v2 catalog release with fileTypes is suggested for an unknown file,
  View plugin opens its preview without installing, and Don't suggest is
  remembered.
The marketplace spec's Git fixture helpers move to a shared helper.

* test(plugins): say why the file-type suggestion e2e needs a longer timeout

* test(editor): install Monaco's cancellation guard in the owner view-state test

The test mocks monaco-setup, which skips the guard production installs, so
a disposed editor's pending word highlight surfaced as two unhandled
"Canceled" rejections once plugin syntaxes added a render. The run
passed but reported errors.

* plugins: plugin app themes recolor the code editor

When a plugin app theme is active, every Monaco surface (editor, diffs,
notebook cells, code excerpts, automation prompt) uses a theme derived
from the colors the document is actually painted with: editor background,
text, line numbers, cursor, selection, current line and widgets. Syntax
colors stay the built-in light or dark ones; if text would fall below
4.5:1 or a syntax color below 3:1 on the plugin's editor background, the
editor keeps its built-in theme. Colors are resolved to hex through the
browser's own parser, so color-mix() and other CSS color forms work. The
theme is redefined when the plugin updates or the mode changes, and the
editor returns to the built-in theme when the plugin theme is deselected,
disabled or removed.

* refactor(plugins): one module per contribution point and a single points table

The manifest's contributes object, the execution-profile derivation and the
catalog type list now read one PLUGIN_CONTRIBUTION_POINTS table instead of
three hand-kept lists.

* refactor(plugins): express id and path rules as patterns

Plugin ids and relative paths now validate with a regex first, so a generated
JSON Schema can carry them as `pattern`. The predicate only adds what a
pattern cannot (reserved names, Windows device names), and a bad value
reports one issue instead of two.

* plugins: document every manifest and catalog field in the schema

Each field now carries a description, examples and the Orca release that
first reads it, and each contribution item has a named fragment id. Manifest
v2 accepts an optional $schema key for editor completion, and the catalog
gains an authoring schema built from the same field definitions.

* plugins: generate published JSON Schemas and author references from zod

pnpm generate:plugin-schemas writes schemas/plugins/orca-plugin.schema.json,
schemas/plugins/orca-marketplace.v2.schema.json and the matching reference
pages under docs/reference from the runtime schemas. A test fails when the
committed files drift, when a field lacks a description or since-version,
when a documented example does not validate, and when the schema and the
runtime disagree on bundled manifests or known-bad inputs.

* plugins: report every manifest and catalog mistake with its path

Manifest and catalog validation now lists up to eight distinct problems as
`contributes.themes[0].label: must not be empty` instead of zod's first raw
message, and rewrites parser wording (required, unknown keys, allowed
values, limits) into what the author should change.

* plugins: say which Orca a newer plugin or catalog release needs

A manifest the strict schema rejects is pre-read for engines.orca and
manifestVersion first, so a plugin built for a newer Orca reports
"requires Orca >=X (this is Y)" instead of unknown-key noise. A catalog
release whose minimumOrca is above this build stays listed but is refused
before anything is downloaded.

* plugins: version the published schemas and pin v2 fields to a real release

Schemas move under schemas/plugins/v2/ so a later breaking format can
publish beside them without changing what a pinned URL means, and the
example plugin authors copy carries its own copy for offline completion.
Fields new with manifest v2 cite one constant, PLUGIN_MANIFEST_V2_SINCE
(1.4.232, the release after the latest tag), and a test fails once a tag at
or above it ships without manifest v2. The reference documents format 1 as
legacy.

* plugins: open review when enabling a plugin that was never approved

Dismissing the review dialog records keep-disabled without consent. Turning
the plugin's switch back on then only cleared the disable, so it returned
to "Needs review" with no dialog. Enabling now opens the review whenever
the result is still pending.

* plugins: release details for the plugin detail page

Catalog releases may carry a license, repository and homepage (https only),
an updatedAt date and a short plain-text changelog; an invalid value is
dropped so it never hides the plugin. Manifests may declare license and
homepage. Marketplace listings flatten name, version and those details, and
installed plugins report license, homepage, repository and the version a
rollback would restore.

The JSON lint-staged step no longer fails a commit whose only staged JSON
files are generated schemas the formatter ignores.

* plugins: fold in the schema review

- Parity runs both ways: documented examples also pass their zod field,
  bundled manifests pass both the schema and the runtime, and manifest and
  catalog negatives (traversal, oversize, wrong types, duplicates) fail in
  both. A table lists every rule only the runtime can enforce, and each is
  named in the description of the field it constrains.
- The published files carry the format in their names
  (orca-plugin.v2.schema.json, orca-marketplace.v2.schema.json); $id stays
  on main until the v2 release, and a test demands the release tag once it
  exists.
- The v1 since-value is 1.4.160, the first release that read plugins, and a
  test pins it to the tags. engines.orca documents that format 2 needs at
  least the release that ships it.
- v2 catalog sources strip unknown keys like the rest of the entry, and the
  v2 runtime schema is derived from the one authoring definition, whose
  plugins field documents the duplicate-id rule. Contribution types and
  categories declare uniqueItems; fileTypes items come from the file-type
  schema itself.
- Issue messages no longer depend on zod's English: a missing field is told
  from the input, and every pattern carries its own message. Unknown-key
  lists show five keys plus a count, and each line is capped at 200
  characters.
- A prerelease host (1.5.0-rc.1) now ranks below its release, as in semver.

* plugins: read the rollback version on demand, not while listing

The plugin list is rebuilt at startup and on every plugin change, and each
installed plugin cost a directory read to find the version a rollback would
restore. That moves to plugins:getRollbackVersion, which the detail page
calls when it shows the action, and a test pins the list build to zero
filesystem calls.

* plugins: v1.4.232 shipped without manifest v2, so v2 fields cite 1.4.233

The release-tag check caught it: v1.4.232 was cut from main before this
stack landed.

* plugins: list installed plugins while the plugin system is off

With the system off the service discovers nothing, so plugins:list returned
an empty list and the settings page could not show what is installed. The
client list now reads the install tree on demand in that state and projects
every plugin as disabled; nothing is activated or started, and startup with
the system off still does no plugin I/O. Older clients that hid the list
while off are unaffected.

Also for the plugin page: plugins:getInstalledManifest returns one installed
plugin's parsed manifest on demand, marketplace installs report the catalog
ref beside its URL, and the listing type declares the release record it
already carried.

* plugins: final schema decisions for catalogs and versioning

Catalog
- The catalog is orca-catalog.json with a required catalogVersion: 2; a
  client that sees a higher version refuses with "update Orca to browse
  this catalog". New clients read it first and fall back to the frozen
  orca-marketplace.json, which export-legacy-plugin-catalog.mjs generates.
- Parsing is tolerant at every level: unknown keys are stripped, unreadable
  entries and releases are skipped, over-limit lists are trimmed, a repeated
  id keeps the first copy, invalid display-only fields are dropped.
  pnpm validate:plugin reports each of those as an error for authors and CI.
- Entries list releases newest first; Orca offers the first whose
  minimumOrca it meets, and shows "Needs Orca X" only when none fits.

Versioning
- engines.orca is the only gate and must cover the newest since-version of
  every field and enum value used (capabilities, events and built-in actions
  now carry per-value since tables). The check is part of the v2 schema, so
  validation, catalog checks and installs all enforce it.
- pluginApi is optional in format 2 (defaults to 1) and refused on data and
  instructional plugins. Format 2 rejects unknown keys at every level,
  including engines, author and every contribution item.
- contributes.agents is reserved for coding-agent definitions and must be
  empty in format 2 until their format exists.
- Plugins are gated against the newer of the app version and the newest
  since this build defines, so development and local builds, whose package
  version only moves at release, read the fields they introduce.

The policy is written into the generated references and pinned by tests.

* plugins: one-click install for content-only plugins, and the verified profile on the list

installMarketplacePlugin takes an optional enableIfData. Main then turns the
plugin on in the same step only when the installed, verified manifest is
execution: data, approving exactly the fingerprint it just installed through
the normal consent path; anything else comes back needsReview: true, whatever
the renderer sends. Older hosts ignore the field and leave the plugin pending.

Installed plugins now report their verified execution profile, so the plugin
page never infers "content only" from the absence of a worker.

* plugins: images and contribution summaries for plugin pages

Manifest format 2 gains cover and screenshots (alt text required), and its
icon must be a square PNG or WebP. Images are checked at install from their
headers only: static PNG, JPEG or WebP, per-kind pixel and byte limits, at
most 4096 pixels a side and 8 MB in total, so nothing is ever decoded to
decide trust.

Catalog releases may carry images stored in the catalog repository and
pinned by SHA-256, a summary of what they add, and the capabilities they ask
for, so a listing can show the plugin completely before anything is
downloaded. All three are dropped when unreadable; the summary and
capabilities must match the manifest exactly at install. pnpm validate:plugin
checks each image's hash, size and format against the repository files.

* plugins: serve plugin and catalog images from verified bytes

Images reach the renderer only as orca-plugin-image URLs, which main serves
with nosniff and the type it read from the header:
- catalog images are read from the catalog repository at the fetched
  commit, kept only when their hash, size and format match, and cached by
  SHA-256; a listing references only those;
- preview images are cached by hash before the staging tree is deleted;
- installed plugins' images are named by plugin, install hash and slot and
  re-checked on every read, so listing plugins still reads no files.

Listings, previews and installed plugins carry an images field the plugin
page can render without fetching anything else.

* plugins: catalog theme palettes, and one-click never approves a dev override

Catalog releases may list up to four theme palettes (hex colors only) so a
listing can draw theme cards before install; they are display-only and
dropped when invalid. One-click enable now also refuses when a development
override shadows the installed plugin, since those are not the bytes the
user just chose.

* plugins: one-click approves only the fingerprint the user saw

installMarketplacePlugin takes an optional expectedFingerprint, the
preview's consentFingerprint. With enableIfData, main now approves only
when the installed fingerprint equals it, so bytes that changed between
preview and install stay pending for review instead of being enabled. The
release commit is already pinned by resolvedCommit. Older hosts ignore the
field.

* plugins: gate manifest v2 on 1.4.242 and publish schemas under docs/schemas

v1.4.240 shipped without manifest v2, so the first release that reads it is 1.4.242.
The root-directory guard refuses a new top-level schemas/ folder.

* plugins: read unvalidated schema input through a typed object guard and break the image validation import cycle

* plugins: one place decides an installed plugin's status word

Pull the status ranking, label and enabled/needs-review checks out of the settings row so the
coming list rows and detail page read the same words. No visual change.

* plugins: one list entry shape for installed and catalog plugins

PluginListEntry joins an installed plugin to its catalog listing (or stands for a listing alone),
and matchesPluginSearch is the one search both lists use. No visual change.

* plugins: usePluginCatalog owns catalog loading, refresh and preview

Sources, listings, refresh, the install/update preview and the outside preview request move out
of PluginMarketplaceBrowser into one hook the coming list and detail page share. The browser
keeps only its view state. No visual change.

* plugins: one plugin list with fixed slots, a plugin page, and inline states

- Installed plugins first, then one group per catalog (verified/unverified · owner or URL), capped
  at 6 rows with Show N more; every plugin appears once
- Rows: the name opens the plugin page and covers the row; status, action, switch and ⋯ are
  fixed-width sibling slots in Tab order. Update {version} is the action when a catalog has a newer
  release; Stopped offers Restart with a red mark on the switch; Blocked offers Remove…
- Plugin page in the same pane: what it adds (commands with platform shortcut chips, panels,
  recipes, contribution summary), what it can access, logs, and info. Back link, Alt+← or ⌘[
  return focus to the row
- States inside the list: skeletons, a catalog that fails only affects its own group, no results
  with Clear search, nothing installed. Plugins off keeps the installed list visible, dimmed
- Add plugin ▾ (folder, Git URL) replaces Install plugin; Plugin system stays at the top
- Catalog listings now expose their release (name, version, minimum Orca, types) to the renderer

* plugins: filter by category and catalog, one review for Update all, Needs newer Orca

- Filter popover (category, catalog) next to search; search and filters show every match
- Update all opens one review: updates that ask for new access are expanded with what's new, the
  rest are one line, any can be unticked; installs run one at a time
- A catalog release that needs a newer Orca says so and disables Install, in the row and the page
- Installing from a plugin page shows a review with only access and source, since the page just
  showed what it adds

* plugins: plugin page uses the manifest name; Off status word; official isn't a category

* plugins: catalog sources live in the Add plugin menu

* plugins: search takes its own row when the list is narrow

* plugins: release notes, license and links on the plugin page; roll back names its version

* plugins: fix list review findings

Plugin page owns the pane (back link and plugin header only); Adds to Orca names
every contribution; access never reads as 'nothing' when the check failed; filter
kinds have readable names; Update all is reachable, fetches once per open and
routes updates needing review; installed plugins join their own catalog;
attention-first order; paused list while plugins are off.

* plugins: readable category names on the plugin page

* plugins: Update all keeps its set; reviews queue; updates only from the plugin's own catalog

* plugins: name what an off or unreviewed plugin adds from its manifest; join catalogs by url and ref

* plugins: Needs Orca checks use the version plugins are gated against

Main gates plugins on pluginHostVersion, the newer of the app version and the
newest since-version the build reads. The plugin page compared releases to the
raw app version, so a development or local build showed Needs Orca for plugins
it would install.

* test(plugins): e2e journeys follow the redesigned plugin list

The page offers "Turn on plugins" while the system is off, catalog rows share
the installed rows' data-plugin-key, the review action is "Review", and an
enabled plugin reads "On".

* plugins: describe access from the verified execution profile, never from "no worker"

A commands-and-shortcuts plugin's page said "Content only. It has no code that runs."
because the page treated "no worker" as "no code". A worker plugin without capabilities
read "runs in a sandbox". The page now classifies from the catalog release's execution
profile (install refuses bytes whose manifest differs), the fetched manifest, or the
installed manifest's profile from the plugin list's `execution` field. Unknown stays
"Reading…" or "Couldn't read", never content only. The review dialog uses the same
profile, and an older host that omits it is reviewed as instructional.

* fix(plugins): read revision-scoped syntax failures on detail pages

* plugins: use the Blocks glyph for icon themes in the plugin page

* plugins: update latest-value refs after render in the update dialog and page navigation

* i18n: retire translated entries for plugin strings this change rewords or removes

* plugins: one action to install, inline consent, and a picture-first catalog

Settings > Plugins no longer opens a confirm dialog and then a review dialog.
- Content-only catalog plugins install from their card or page in one click; main
  turns them on only when the verified manifest is data. Undo stays in place for 10s.
- Plugins that run code or instructions show "If you install <Name>, it can:" on the
  host security surface above one Install & enable, which stays off until the access
  rows are read, and approves exactly the fingerprint shown. Review of an installed
  plugin happens on its page too (Enable / Keep off); Remove stays in its menu.
- The catalog is a grid of cards drawn from what each plugin is (its theme colours once
  installed, a code sample, keycaps), with the name as a stretched link and one action.
- The page pins the listing it opened, says "Newer version available" instead of
  swapping rows, and previews are prefetched only on hover intent, capped per visit.
- The plugin system switch, file suggestions, catalog sources and development folders
  move into a header menu; amber text uses a contrast-safe token.

* plugins: give the reworded no-access line its own key so English shows the new copy

* plugins: pin installs to the release shown, warn by what a plugin carries, disclose replaced shortcuts

- The page pins the listing it opened; a preview of any other release (commit, catalog
  commit or execution profile) leaves Install & enable off and shows "Newer version
  available". One-click installs from a card refuse the same mismatch.
- Warnings are the union of the verified profile and the contributions: a panel that
  also adds commands gets the instructions warning, never "bounded" or "own files only";
  unknown contributions over-warn. Access rows follow the same rule.
- Shortcuts that replace one of yours are named in the consent before install, from the
  manifest or the catalog summary; the shortcut list itself stays in Adds to Orca.
- Pages paint from the catalog summaries, capabilities and theme palettes, and show the
  cover (or the card's art) as a hero, with screenshots in a strip and lightbox.
- Installing fills a reserved status line and keeps its action row (Undo, Use now), so
  the page doesn't move; the settings scroller keeps a stable gutter.
- Row art names the plugin type instead of initials, rows drop the "Needs review" word,
  the content hash moves into the menu, and card art is tinted and coloured.

* plugins: catalog covers on cards

* plugins: Review reopens its page after Back; hero art and stable status line follow what an installed plugin carries

* test(plugins): e2e journeys review and install on the plugin page, with no dialogs

* plugins: install exactly the release shown, classify from the full manifest, polish the page

Code review:
- A page or card arms Install only when the preview matches everything the listing showed
  (catalog commit, release commit, content hash, execution); any change, including a
  catalog-only refresh, reads "Newer version available". One-click installs send the
  preview's fingerprint as expectedFingerprint, so main turns on only those bytes.
- Install stays off while a catalog refresh is in flight.
- Trust uses the full verified preview manifest when present; a capped catalog summary
  counts as unknown and gets the stronger warning.
- The installed-manifest read is keyed by plugin and installed revision, so a same-key
  update never pairs old contents with the new fingerprint.
- A failed Undo shows its error on the card or page and keeps Undo to retry.
- The security-surface test now covers the inline consent instead of the removed dialog.

Visual:
- Lightbox: close button clear of the image, "1 of N" with previous/next and arrow keys,
  content scrolls on short windows, focus returns to the opening thumbnail.
- Keycap art shows the plugin's own first shortcut, larger, with the platform's keys.
- The hero fills the content column; a screenshot identical to the cover isn't repeated.
- "Change in Shortcuts" is a link; the chip says "Adds commands" (or "Runs a worker",
  "Adds a panel", "Content only"); the disabled button says "Reading what it can access…".
- A content-only plugin's action row keeps fixed slots, so its note doesn't move on install.

* plugins: keep the decision above the fold, route suggestions to the page, final review fixes

- The page shows the decision (Install, the consent, Enable) before the picture, and the
  picture is capped and cropped from its centre, so the button is in view at 1280x800 and
  900x700. Nothing grows after first paint.
- The consent gives no access verdict while the plugin is still being read.
- A file-type suggestion opens the plugin's page instead of the old install dialog; that
  dialog now serves updates only.
- A theme plugin reserves its "Use now" line from first paint; the action slot and its
  button share one width.
- The file-suggestion toggle reports a failed save; a failed Undo is reported once, on
  the card or page that offered it.
- Cards show a plain trust chip ("Content only", "Adds commands"); the local and bundled
  group reads "Bundled & local".
- Install error copy moves into the protected plugin namespace, and the new consent
  components join the language-pack security pins.
- Dead code goes: the keybinding consent preview (its conflict helper stays), its strings,
  trustRunsNothing and the access-only preview plumbing.

* test(plugins): the file-type suggestion opens the plugin page, not a dialog

* plugins: the install decision sits top-right beside the name; no plugin on/off switch in the UI

- Plugin page: Install (content only), Install & enable (disabled until its access is read),
  Enable / Keep off (installed, waiting for review) and the installed switch share one slot
  top-right of the title. The consent sits directly under the header with no button of its
  own; the "Can't reach anything outside its own files" note sits on the line under the
  publisher, which every state reserves, so nothing moves between states.
- Removed "Turn off plugins" from the ⋯ menu and the "Turn on plugins" button from the off
  state. The setting and all gating stay: off now comes only from policy, and the page says
  so without a switch.
- e2e journeys turn plugins on through the setting and use the page's one button.

* plugins: restore Turn on plugins, drop Undo, add Remove to installed card menu

Plugins stay opt-in, so the off state keeps its Turn on button; only the
menu's Turn off item stays removed. Installed catalog cards get a sibling
more-actions menu with Remove (immediate, inline error) and Turn off/on,
inside a fixed-width action slot so the footer doesn't shift.

* test(plugins): installing a theme plugin never changes the active theme

* fix(plugins): confirm card removal before deleting data

* fix(plugins): pause review queue when leaving plugin page

* fix(plugins): disclose execution changes before updating

* refactor(plugins): share single and bulk update review

* refactor(plugins): share installed card and row actions

* test(plugins): cover confirmed removal and shared update review

* feat(plugins): offer icon and terminal themes after install

* test(plugins): capture settled review dialog screenshots

* fix(plugins): make installed contributions discoverable and actionable

* fix(plugins): label card-install toast Try it and mark terminal-theme Use now buttons

The toast action opens the plugin page rather than applying a theme, so it
always says Try it. Terminal themes often share their app themes' names, so
their Use now buttons carry a terminal icon. Update the update-review e2e for
the Use now accessible names.

* fix(plugins): apply a theme plugin's themes when the user installs or enables it

Installing (card or page) or enabling a plugin that adds themes now applies its
app theme (matching the current light/dark mode), icon theme and terminal theme
at once, with a 'Now using' toast whose Switch back restores the previous
selections. Updates, startup discovery and catalog refreshes never apply.

Co-Authored-By: Claude <noreply@anthropic.com>

* plugins: call toast actions through a typed button in tests; key card art by stable names

* i18n(plugins): drop retired plugin strings from the Portuguese catalog

* test(plugins): say why the update review e2e needs a longer timeout

* i18n: retire translated entries for plugin strings this change rewords or removes

* refactor(plugins): organize main plugin files by responsibility

* refactor(plugins): group shared plugin schemas and contracts

* refactor(plugins): organize renderer plugin surfaces

* plugins: keep the plugin service under the line limit after the folder moves

* refactor(plugins): move the JSON object guard beside the manifest schemas

* test(plugins): move the flaky-test baseline entries with the worker tests they count

* test(plugins): point the panel lifetime test at the reorganized plugin folders

* plugins: reserve skills, mcpServers, taskSources, connections and settings contribution keys

Format 2 accepts each key only as an empty list, so the points can ship later as additive
minors without a new manifest format.

* plugins: freeze host API v0 with a generated reference and refuse unknown pluginApi majors

The method table now carries a summary and an optional deprecatedSince, and renders
docs/reference/plugin-host-api.md with the stability and deprecation policy. A manifest
naming a newer host API major fails with "update Orca" instead of "must be 1".

* plugins: name the bad release in catalog validation and re-ask on any execution profile change

validate:plugin now reports each unreadable release with its reason (for example a tag
instead of a full commit SHA) rather than a count. The consent fingerprint includes a
declared non-data execution profile, so a data pack that becomes instructional no longer
keeps its one-click approval. Tests pin that catalogs drop non-SHA releases, that a
repository cannot load or approve a plugin it carries, and that added capabilities or a
changed profile always need re-approval.

* plugins: add the author guide, theme token list, v2 examples and a printed catalog release entry

docs/reference/plugin-authoring.md walks write, load a development folder, validate and
publish. The theme token table is generated from the public token set and its fallbacks.
hello-orca and hostile-panel move to manifest format 2. validate:plugin on a plugin
repository root materializes HEAD the way install does and prints the release entry to
paste into orca-catalog.json.

* plugins: name the host API type renderer and reserved contribution fields for their role

* plugins: gate the bundled example plugins on 1.4.240

* plugins: reserve orca, orca-*, official, anthropic, openai and github publishers

* plugins: rename storage/secrets to storage:own/secrets:own and gate events by domain capability

events:subscribe and the events.subscribe host method are gone; workspace events need
workspace:read and agent status needs the new agents:read. Format 1 manifests keep their old
names, rewritten on parse.

* plugins: move samples and fixtures off the now-reserved orca-samples publisher

* plugins: name the public surface workspace, not worktree

Events become workspace.created/removed with workspaceId and a kind field, agent status
carries workspaceId, and command context/keybinding when use workspace. Format 1 names are
rewritten on parse, v1 workers still receive worktree.* handlers, and clients treat any
non-global context from older hosts as workspace-scoped.

* plugins: rename the instructional execution profile to declarative

* plugins: move agents into the reserved contribution keys

Format 2 keeps contributes.agents empty like every reserved point; format 1 entries, which
Orca never read, are dropped on parse so they stay inert and bind no content.

* plugins: skip unknown theme tokens and icon slots at runtime, refuse them in validate:plugin

A token or slot added in a later release no longer fails the whole theme on older Orca.
validate:plugin now parses each theme, terminal theme and icon theme strictly. App themes
accept an optional schemaVersion 1.

* plugins: inject panel design tokens as --orca-* custom properties

* plugins: accept a closed, documented list of panel icons

Format 1 panels keep any icon name: a known one in either Lucide spelling is canonicalized,
an unknown one is dropped and shows plug, as before.

* plugins: make catalog release types optional and derive them

Every required release field is permanent for old clients, and types repeat what contributes
and execution already say. Printed release entries no longer include it.

* plugins: strip unknown kill-list and consent fields instead of refusing them

A strict kill list froze old clients on their cached list as soon as a later list added a
field, silently stopping revocation; a strict consent request would refuse a newer client's
optional review field.

* plugins: check only the shape of a panel action in the renderer

The host's capability gate already refuses methods panels may not call. Filtering against the
client's own method list blocked every panel method a newer host added on older clients,
including remote and SSH ones.

* plugins: read a leading >=x.y.z from engines.orca leniently in the compatibility pre-read

* plugins: derive content-bound approval from the contribution policy table

Each point now declares bindsContent, replacing the hand-kept list in the consent
fingerprint, so a new point such as mcpServers cannot swap bytes under an old approval by
being forgotten.

* plugins: name the host API "Host API 1" and add a stability section to the author guide

The guide now lists what is stable and what is experimental: the agent status payload, the
worker process model, language-pack keys, validate:plugin output beyond its exit codes,
settings:own and manifest format 1. Secrets copy now says encrypted file, not OS keychain.

* plugins: make the hello-orca sample speak in plain words

The worker command returns a sentence instead of a JSON object, and the panel labels
terminals by position instead of their opaque ids.

* plugins: give plugins short branch names, not refs/heads/...

* plugins: panels can no longer type into terminals; workers only (D3)

terminal.sendText with Enter runs shell commands locally or over SSH, which
the sandboxed-panel tier's "limited" consent copy did not convey. The
capability copy now says so, and the sample panel no longer types.

* plugins: put full-access worker plugins behind a second setting, off by default (D2)

Trusted workers are unsandboxed Node, so the plugin system switch alone no
longer lets them run. With "Allow plugins that run code with full access"
off, main refuses to install a full-access catalog release, approve or enable
one, and keeps installed workers stopped with an error naming the setting.
The catalog card says "Full access · turned off in settings" and links to
the switch; the worker chip now reads "Full access".

* plugins: bind dev-folder approval to the folder it was given for

A dev plugin's consent fingerprint now includes the folder's canonical path,
so approving one dev folder no longer approves another folder whose
manifest claims the same key and capabilities. The development dialog now
says edits inside an approved folder run without another review.

* plugins: close WebRTC as an outbound channel from panels

CSP has no WebRTC directive, so a panel could reach a STUN/TURN server it
controls despite connect-src 'none'. The panel shell now locks
RTCPeerConnection to undefined before plugin code runs, and the main window
that hosts panels gets the same WebRTC IP-handling policy as doc preview
guests. DNS-prefetch hints remain the residual doc preview already accepts.

* plugins: stop language packs relabelling trust copy outside plugin settings

Protection covered only plugin copy under settings, so a content-only pack
could rename the catalog's Official badge, the file-type suggestion banner,
or swap Allow and Deny on agent approvals. Those components are now
protected whole, the same way the plugin security dialogs are.

* plugins: refuse Git URLs with a dash-led user or host, or a '..' host

ssh reads a leading '-' as an option, and the scp form is not normalized, so
'stablyai/../evil' read as the official owner. The shared URL check used by
the catalog schema, validate:plugin and every install path now refuses these,
and provenance parsing refuses '..' path segments.

* plugins: flag an update whose catalog moved it to another repository

An update whose catalog now points at a different repository showed the new
URL without saying it changed. The update review now names the old and new
repository and leaves that update unticked, so it installs only once the
user opts in.

* plugins: sample worker returns a readable toast; sample panel shows no internal ids

Hello Orca's ping returned raw JSON, which the command toast printed as-is,
and its panel showed the full ref name. It now returns a sentence and shows
the plain branch name.

* plugins: move runtime policy checks and catalog types out of files over the line limit

* plugins: refresh the plugin safety list every 6 hours while plugins are on

Revocations previously reached a session only at launch or when plugins were
turned on, so a long-running Orca kept running a revoked plugin.

* plugins: main sends an update from a moved repository to review

The update dialog left such updates unticked, but main still accepted a
one-click approval for them. Main now compares the installed and new
repository itself and always answers needsReview.

* plugins: label the plugin system Alpha instead of Experimental

Settings badge, nav entry badge, search/description copy in all locales,
and plugin docs' stability sections. Setting keys are unchanged.

* plugins: trim the Plugin settings menu to titles and a short full-access hint

* plugins: reserve statusBarItems and sidebarCards and list planned contribution points

Both keys are rejected when non-empty, like the other reserved points. The
manifest and authoring references gain a Planned contribution points list
built from one table. Also moves the Alpha note on the host API page into its
generator so the schema/doc parity test passes again.

* plugins: an update that asks for nothing new keeps its approval

Before, any catalog update to a declarative, sandboxed-ui or trusted-worker
plugin changed its consent fingerprint (it binds the package bytes), so the
plugin turned off and waited for review after every Update. Only content-only
(data) plugins kept running.

Main now reads the approved version before installing and, when the update
comes from the same repository and adds no capability, no higher execution
profile, and no new command, shortcut, panel, event or VM recipe, records the
new fingerprint so the plugin stays on. Anything that grows access, a moved
repository, a development override or a plugin that was never approved still
waits for review, and the full-access switch still gates trusted workers.
The update preview now carries what grew, so the Update dialog lists new
shortcuts, commands, panels and events as well as capabilities. The consent
copy says updates that don't ask for more access install without asking.

* plugins: match the full-access test to the shorter setting name

* plugins: update catalog plugins automatically when an update asks for nothing new

Adds "Update plugins automatically" to Plugin settings, on by default. The
setting is pluginUpdatePolicy ('auto' | 'manual') so later policies can be
added without a migration.

At startup (after a short delay) and every 12 hours, packaged Orca refreshes
the catalogs and, for each enabled, approved catalog plugin with a newer
release, installs it in place when it comes from the same repository and
grows no access; the approval carries over and the worker reloads. A release
from a non-official catalog installs only after it has been listed for 12
hours, timed from when this Orca first saw it (stored locally because
catalogs carry no listed-at time), so takedowns can land first. An update
that grows access or moved repository is not installed: the current version
keeps running and the plugin shows "Update needs review", whose Review
update button opens the existing update review with what is new. Plugins
updated this way show a quiet "Updated to vX" chip, and each pass shows at
most one toast. Manual Update and Update all are unchanged.

* plugins: clear code-quality findings in the update approval and auto-update code

* plugins: give an app theme that leaves --sidebar unset a visible sidebar edge

Orca draws no line between the left sidebar and the main area; the sidebar's
slightly different surface is the edge. A theme that omitted --sidebar (like
the Dusk sample) had it filled from --background, so the sidebar matched the
main area and the edge vanished. An unset --sidebar is now mixed 8% toward
the theme's foreground. The in-repo sample theme sets --sidebar explicitly
and the theme smoke test checks it reaches the sidebar.

* plugins: treat an unknown update policy as manual, not auto

* plugins: auto-update only pinned releases, and only the release that was held

* plugins: start the auto-update hold clock before the catalog refresh, and check every 6 hours

* plugins: refresh the kill list right before auto-installing, and skip the install if that fails

* plugins: refuse an unpinned full-access install while full access is off

* plugins: an automatic update re-checks inside the install lock that the plugin is still installed

* plugins: compare the full repository path, so a GitLab subgroup move counts as moved

* plugins: time out the kill-list fetch and retry 15 minutes after a failure

* plugins: drop one unreadable kill-list entry instead of the whole list

* plugins: keep language packs off the orca.yaml and notebook trust prompts, certificate warning and site permission prompt

* plugins: compare whole capability declarations on update, and cover swapped command actions and shortcuts

* plugins: log automatic updates so headless and remote hosts record them

* plugins: say that automatic updates can change what a full-access plugin does

* plugins: pin and document the nested-frame WebRTC residual in plugin panels

* plugins: delete the unused PluginReviewBody

* plugins: drop the dead repositoryMoved parameter and the predecessor wrapper

* plugins: declare the plugin policy settings type once

* plugins: show what main found an update changes instead of recomputing it in the renderer

* plugins: move update approval and update policy into the catalog folders

* test(plugins): expect the workspace.removed event name in the lifecycle test

* plugins: rename structural helpers for their role, type the update identity read, and give plugin system settings their own group

* i18n(plugins): drop retired plugin strings from the Portuguese catalog

* plugins: build the hello-orca status line with a template string

* plugins: install catalog plugins from a directory of a shared repository

A catalog source may name a 'path' inside its repository. Orca downloads the
pinned commit as before but materializes only that subtree, so contentHash
covers the plugin's own files and sibling commits never change it. A changed
path counts as a moved repository and needs review. validate:plugin accepts a
subdirectory and prints its source path; the frozen v1 export skips such
entries.

* plugins: offer the full access setting on a full-access plugin's page

The page offered Install & enable even with full access off, then failed with a generic error.
It now shows the card's 'Turn on in settings' action, names the real reason when main refuses,
and drops a stale install error once the full access setting changes.

* plugins: say a content-only plugin couldn't load its files, not that it crashed

Language packs and other data plugins have no code or logs, so a load failure now reads
"Couldn't load its files" on the page and the row instead of "Its code kept crashing".

* plugins: leave Settings before running a built-in action from a plugin page

Built-in actions such as Open Board and Go Back are disabled while Settings covers the app,
so Run on the plugin page always failed. Run now closes Settings, commits that synchronously
so the dispatcher sees the app view, then runs the action.

* plugins: refuse terminal:send on panel plugins and never list it on panel approvals

Panels get no host method that types into a terminal, yet a panel's approval listed
"Type and run commands in your terminals". Format 2 manifests now fail validation when a
profile below trusted-worker asks for terminal:send, and approvals, the access section and
update review drop that line for format 1 panels that still declare it.

* plugins: say "catalog" in user-facing copy, not "marketplace"

The menu said "Catalog sources…" but the dialog it opens said "Marketplace sources" and
asked for orca-marketplace.json. Plugin copy now says catalog in every locale, names
orca-catalog.json (still accepting the older file), and the author guide matches the UI.
Internal identifiers and settings keys are unchanged.

* test(plugins): run a built-in action from its plugin page end to end

* plugins: concise copy pass across plugin UI

Update review shows one line per plugin and only what's new for updates
that need approval, with the rest behind a See all access disclosure.
Consent, Can access, dialogs, settings menu and empty states drop restated
titles and filler; safety facts (new access, full access = files/network)
stay visible where the user decides.

* plugins: keep See all access on Button and Collapsible primitives' own spacing

* plugins: separate light and dark theme slots so packs pair and mix

Plugin app themes now fill a Light theme or Dark theme slot by their base. Orca paints the
slot for its current mode, so under System it swaps live with the OS. Appearance shows both
pickers grouped by plugin; install fills the matching slot(s) and Switch back restores them.
A legacy single selection migrates into its base's slot.

* plugins: installed carousel and card grid, category badges, In use state; load language packs that cover protected copy

* fix(theme): let plugin themes colour menus and dialogs

Menus, popovers, selects, hover cards, sheets and dialogs painted a fixed
glass tint (dialogs rgb 23,23,23 in dark mode), so a plugin theme's
--popover was ignored and a true-black theme still showed grey dialogs.
The tint now lives in --floating-surface / --dialog-surface, which a
plugin theme replaces with its --popover.

* plugins: name the app-theme slot a pack fills, not its terminal colours

* plugins: page the installed row by whole cards; e2e follows the card menu

* plugins: hide the installed row's scrollbar and draw kind art instead of initials

The installed carousel pages with its arrows and snap points; the native
scrollbar under it only added a stray bar. Format 1 catalog listings (the
official catalog until catalog v2 publishes) carry no release types or
images, so their cards fell back to a two-letter monogram. Art now reads
the kind from catalog categories, and the last fallback is a plugin glyph.

* plugins: reserve official publishers to Orca's catalog and pin installs to their catalog

Reserved publishers (stablyai, official, anthropic, openai, github, orca*; case-insensitive)
and orca-* ids now install only from Orca's official catalog: a Git URL, local folder or any
other catalog is refused, and a third-party catalog listing one is refused whole. The rule lives
in one import-free module the directory indexer can vendor; validate:plugin uses it too and
warns when publisher differs from the origin remote's owner.

A marketplace install never replaces one recorded from a different catalog; switching means
remove and reinstall. Docs state the publisher == repository owner policy (enforced by the
official and community catalog CI, not a new runtime field).

* plugins: label trust by catalog source and note other catalogs listing an install

The chip beside a plugin now says where trust comes from: the Official mark for Orca's catalog,
otherwise the catalog name and the owner parsed from its Git URL (github.com/acme,
gitlab.com/acme/tools), never manifest text; Dev · unverified for development folders and
Unverified for Git URL installs. An installed plugin's page says which other catalogs list it.

* test(plugins): give update review's Git-backed steps a loaded-runner timeout

Install and update review each fetch a plugin over Git. On a loaded machine that
took 25-30s with or without the publisher checks, past the 10s expect default,
so the test read as stalled on Installing…

* test(plugins): give the row-art fixture the entry's catalog source fields

* test(plugins): say why the panel WebRTC test needs a longer timeout

* feat(plugins): dismissible Alpha notice on the Plugins settings page

Shows "Alpha: The plugin interface may change. Expect nuances." under the
Plugins title until dismissed; dismissal persists in UI state
(pluginsAlphaNoticeDismissed). Localized in all locales, with unit and e2e
coverage for show-by-default and dismiss-across-reload.

Co-Authored-By: Claude <noreply@anthropic.com>

* plugins: stop shipping Orca Navigation Shortcuts and retire stale bundled installs

Removes the bundled copy and its fallback-catalog listing. On startup, bundled
bootstrap now uninstalls any previously bundled plugin this release no longer
ships, so existing users don't keep a protected entry they cannot remove.
Removal code moves to plugin-removal.ts to keep plugin-install.ts under the
line limit.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(plugins): concurrent saves of one image no longer fail the update review

Each save now writes its own uniquely named temp file and treats a rename
failure as success when a concurrent writer already produced the same
content-addressed file. A cache failure now only drops the image from the
preview or catalog refresh instead of failing it.

* fix(plugins): an update keeps unchanged machine recipes off the review's New list

Recipes always counted as new because their commands live in files the
manifest diff cannot see. Main now reads what each recipe can run (id,
checkout mode, commands) from the installed tree and the preview's staging
tree, and only a recipe that runs something new asks for review. An
unreadable file still counts as new.

* plugins: app themes color code in the editor via VS Code-shaped tokenColors

Theme files gain optional tokenColors (TextMate scope -> foreground/fontStyle).
The derived Monaco theme maps them onto Monaco's token names, the plugin-syntax
legend and raw TextMate scopes; unreadable colors are dropped.

* test(plugins): theme smoke checks tokenColors color the editor and revert

* i18n: retire translated entries for plugin strings this change rewords or removes

* plugins: add the custom editors contribution point

A plugin can now declare contributes.customEditors to open files of given
types in the editor tab, beside Source. The editor runs in the same
sandboxed, no-network panel frame and only exchanges the document's text
over a scoped postMessage channel; Orca keeps the buffer, dirty state,
saving, external-change banner, read-only tabs and SSH files.

The Editor | Source toggle and the 'first scene after load is the
baseline' approach follow community PR #27413.

Co-authored-by: Roughriver74 <115543510+Roughriver74@users.noreply.github.com>

* plugins: list custom-editor plugins with panels in the plugin list

* plugins: name an editor-only plugin's access 'Adds an editor'

* plugins: keep the Editors category label for custom-editor plugins

* plugins: restyle plugin frames in place and keep file paths out of editor frame names

A theme change used to remount panel and custom editor frames, which reloaded
the plugin and lost its state (Excalidraw undo, unsent edits). The host now posts
orca-panel-theme and the shell updates the --orca-* tokens and the html class in
place. Custom editor frames are named with a random id: window.name is readable
inside the frame, and the old name carried the file's absolute path.

* plugins: let custom editors open binary files as bytes (contentKind)

A custom editor can now declare "contentKind": "bytes" (with "readonly": true)
to receive the file's raw bytes instead of its text, for formats such as
spreadsheets. Orca pages the bytes from the host that loaded the tab (local,
SSH or a paired runtime, through the existing stat + readChunk calls, so no
wire change), refuses files over 16 MiB with a "too large to open" message,
re-sends on outside changes, and never applies a bytes editor's edits. The frame
gets a transferred ArrayBuffer and the matched suffix, never a path.

Documents are now versioned: { version, readOnly, content: { kind, ... } }.
Text edits may name a baseVersion; an edit against a replaced document is
refused with an orca-editor-edit-rejected conflict. Text documents keep the
legacy text field and unversioned edits still apply, so existing editors work.

"readonly": true on any editor makes the install review say "Opens .x files in
Orca" instead of "Opens and edits".

* plugins: suggest catalog custom editors for files Orca already opens

The file-type suggestion only offered a plugin when a file opened as plain text
or as the binary placeholder. It now also offers a catalog custom editor for a
file Orca shows with a language (an .excalidraw file opens as JSON): the editor
asks main with the `editor` purpose, which matches only release file types the
listing's contributions name as custom editors, and only when no installed
plugin's editor handles the type. Binary files such as .xlsx keep the existing
path, since release fileTypes already include custom-editor types. The banner
reads "A plugin adds an editor for .x files: Name"; Not now and Don't suggest
work per type as before.

* fix(plugins): an update that lets a custom editor edit, take over opening, or change content kind asks again

* fix(plugins): one unreadable catalog summary item no longer hides the whole summary

* fix(plugins): plugin editors never open Orca-core file types by default, and join Markdown, CSV and notebook view toggles as options

* fix(plugins): catalog releases list editor file types as data instead of parsing the display summary

* fix(plugins): custom editor edits must name their base version; drop the pre-release text field

* fix(plugins): a custom editor edit Orca refuses keeps the tab dirty and blocks the save instead of vanishing

Rate-limited edits are held and applied before the next save; oversized ones are reported to
the frame and the user. Every save, from any trigger, first flushes plugin editor frames with a
bounded wait and refuses to write while a frame holds edits the buffer lacks.

* fix(plugins): run custom editors as network-isolated guests and save only on real key presses

Editors see file contents, so they now run as <webview> guests in their own in-memory partition
instead of iframes of Orca's renderer. The session routes everything through a dead SOCKS proxy
(loopback included), refuses every non-editor request and permission, and each guest gets
disable_non_proxied_udp, no popups and no navigation. An Electron test proves fetch, WebSocket,
images and TURN-over-TCP all fail from the editor and from its srcdoc and javascript: child
frames, while the same page in a plain session reaches every listener. Electron 43 still resolves
<link rel=dns-prefetch> names outside the session; the test pins that residual and the docs say so.

Save chords now come from main's before-input-event on the guest, which page code cannot
synthesize; the shell's keydown forwarder and the orca-editor-keydown message are gone. Guests
also run in their own renderer process.

* fix(plugins): enforce a read-only custom editor in main, not just the tab

Main reads readonly from the installed manifest (bytes editors always count) and drops every
edit such a guest posts before it reaches the renderer. The tab host also refuses them and tells
the editor its document is read-only, even on a writable tab.

* test(plugins): the editor-suggestion catalog fixture lists editorFileTypes

* plugins: custom editors edit and save binary files; drop readonly and the isolated guest

Bytes editors now post `{ baseVersion, content: { kind: 'bytes', bytes } }` full
replacements (a transferable ArrayBuffer). The tab keeps the unsaved bytes beside
the text drafts, marks itself dirty, and the normal save pipeline (frame flush,
Cmd+S, close prompt) writes them: locally through a new `fs:writeFileBytes`
atomic replace guarded by the loaded bytes' SHA-256, and on SSH-served or paired
hosts by uploading chunks with `files.writeBase64Chunk` and swapping them in with
the new capability-negotiated `files.replaceWithUpload`. Older hosts are refused
before anything is written. A stale baseline is a conflict: the edit stays
unsaved and the external-change banner offers Reload or Overwrite. Source stays
the binary placeholder.

Per the product decision, every custom editor may edit: the `readonly` manifest
field, its main-side enforcement and re-approval rule are gone, and an editor
that only views simply never posts edits. Editors run in the normal sandboxed
plugin frame again instead of a network-isolated webview guest; the frame's save
chord is honored only while it carries real user activation. Consent now reads
"Opens and edits .ext files in Orca. It can read the files it opens."

* plugins: reserve agentHarnesses and hooks instead of agents; rename agents:read to agent:read

* plugins: syntaxes and custom editors share one file matcher (extensions with the dot, plus filenames)

* plugins: title for themes and syntaxes, when for commands (format 1 names rewritten on parse)

* plugins: add the editor catalog type for custom editors

* plugins: official means listed by the official catalog, not an orca- id

* plugins: one envelope protocol and window.orca for workers, panels and editors

Workers, panels and custom editors now share one internal envelope
({orca: 1, kind: req|res|evt|cancel}) instead of three message sets.
Frames get a frozen window.orca from the shell (api, host.call with
AbortSignal cancel, host.has, on, and document helpers in editors);
ready/hello carries the api, the surface's methods and the theme.
Editors gain host calls through their own session.

The host API table is keyed by method with typed params/results (no
handler casts), per-method surfaces and a numeric since minor; every
params schema is strict. Errors are {code, message, retryable,
retryAfterMs?} with cancelled/timeout/conflict/not_permitted/too_large
reserved and unknown codes read as action_failed. Event payloads are
typed with a closed agent-status enum, and the list entry type is
shared by main and preload.

* plugins: orca.document.onDidChange hands a late subscriber the current document

Editors that subscribe after their bundle loads (Excalidraw) missed the
document that arrived first and stayed blank.

* plugins: custom editor documents up to 64 MiB, with a large-file note past 16 MiB

The cap covers text and bytes on every route; reads and SSH or paired-host
writes are already chunked, so only the shared cap moves. Past 16 MiB the
tab shows a small, non-blocking note that the editor may take a moment.

* plugins: delete dead props and branches left by removed rows, logs toggle and access retry

* plugins: one status word per state (a crashed plugin reads Stopped everywhere) and drop unused plugin locale keys

* plugins: contribution rows use the manifest point names, so the policy table is the one source of catalog types

* test(plugins): assert the Use now line and page behavior, not Tailwind classes

* i18n: add the custom editor large-file note to the English catalog

* plugins: fold the loose shared plugin modules into contributions, frames, images, capabilities, events, languages, installation and catalog folders

* plugins: custom editor tab icons and contributes.newFiles templates

A custom editor may declare an icon; the tab and its hover card show it and the
editor title instead of "Editor", falling back to the plugin icon. contributes.newFiles
declares untitled file types (title, extension, optional UTF-8 template and icon) offered
in the tab bar + menu, its search, the command palette and the plugin page, created at the
workspace root through the Markdown untitled-file flow. The install toast's Try it creates
one for plugins that declare them.

* test(plugins): hidden e2e for a plugin new file from the + menu and its editor tab icon

* plugins: turn the plugin system on by default with a one-time migration

Saved false from the old default flips on once (pluginSystemDefaultedOn guard);
a later deliberate false is respected. The Turn on plugins off-state is gone,
and the official catalog seed is deferred off the startup path (and skipped
offline), seeding on demand when the marketplace is opened.

* plugins: move the card menu Disable/Enable directly above Remove, without an icon

* plugins: show the Alpha notice with the warning banner tokens

* plugins: show each category's icon in the filter dropdown

* plugins: seed drawn fallback art per plugin so two listings never look identical

* plugins: show manifest covers for catalog releases published without images

The official catalog's newest Midnight, Nord, OLED and Dusk & Ember releases
were regenerated without an images block (release-entry only kept images the
author pasted), so their cards fell back to drawn art. A catalog that keeps its
plugins in its own repository now borrows the manifest's images from the
catalog commit, checked like declared images, and the validator's printed
release entry includes them so regenerated catalogs keep their covers.

* plugins: drop the unused Turn on plugins strings

* settings: find the Plugins page by plugin, mod and extension searches

* plugins: remove the Allow full-access plugins setting

Full-access (trusted-worker) plugins now need only their own per-plugin
approval: main no longer gates install, consent, enablement or runtime on a
global switch, and the settings menu item, the Turn on in settings card and
page states, and the stored key are gone. The consent now says plainly what
full access means. Kill list, official review and publisher checks are kept;
docs note an enterprise block policy may come later.

* i18n: retire translated entries for plugin strings this change rewords or removes

* test(plugins): prune low-value plugin tests

Drop tests that asserted CSS classes, DOM structure, internal call counts,
copy, or duplicated coverage; merge overlapping e2e file-type suggestion
specs. Keep security, consent, data-integrity, compatibility and main
journey coverage.

* test(plugins): stop git auto-maintenance racing marketplace fetch teardown

* test(plugins): import the panel session token schema from the frame bridge

* feat(plugins): open orca://plugins install links at the plugin's page

orca://plugins/<publisher.id>?catalog=<official|community|git URL> opens
Settings -> Plugins at that plugin; it never installs. Plugins off shows
Turn on plugins first, an unknown catalog asks before it is fetched, and a
malformed link only toasts. Reuses the skill-share deep-link path (open-url,
second-instance argv, sidecar handoff) with repeat/burst dedupe, and
registers orca:// on packaged Windows, which NSIS leaves unregistered.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(plugins): show the linked catalog URL on its own line

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(plugins): follow main's moved plugin files and language-pack coverage field

* fix(plugins): keep a plugin disabled through its update and give format 1 event handlers worktreeId

* fix(plugins): keep held editor edits across view switches, ignore stale editor entry reads, share bytes across split panes, and meet the low-evidence lint

* refactor(editor): keep the file content loader under the line limit

* fix(plugins): keep a bytes draft's baseline across reloads, cover plugin new-file menu rows, and move render-time ref writes into layout effects

* refactor(tab-bar): keep the create-menu controller under the line limit

* refactor(tab-bar): keep the create-menu controller under the line limit with the plugin new-file case

* test(plugins): turn plugins back on through settings in the install-link journey

* fix(wsl): import the atomic plugin file writer from its new folder

* fix(plugins): save bytes custom-editor edits from Save-and-close and the overwrite banner

* test(plugins): park the Linux-only bytes editor e2e failure as a follow-up

* fix(plugins): follow main's new language-pack tests to the moved artifact module

* test(i18n): give main's new plugin language-pack fixture its coverage

* plugins: manifest v2 ships in 1.4.245, since 1.4.244 released without it

* fix(i18n): import the plugin language-pack artifact from its folder in the shared i18n instance

* fix(plugins): drop syntax loads that land after dispose and restore terminal overrides on Switch back

---------

Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Roughriver74 <115543510+Roughriver74@users.noreply.github.com>
2026-10-28 20:58:25 -07:00
Neil ca272a78c0 fix(daemon): read the PTY process table without blocking on terminal kill and resize (#34603)
* fix(daemon): read the PTY process table without blocking on terminal kill and resize

The terminal daemon force-killed a session by running ps and reading
/proc/<pid>/stat synchronously, and checked the resize foreground group
with execFileSync('ps'). A slow or wedged ps froze every terminal the
daemon serves for up to the ps timeout on each call.

Both paths now use the async process-table read (runProcess with a
timeout that kills the child). SubprocessHandle.forceKill may return a
promise; the termination controller awaits it so a refused signal still
releases the claim for a retry. A force-kill or SIGWINCH whose root
exits while ps runs signals nothing, since the captured groups could
already be recycled. ConPTY force-kill stays synchronous so dispose()
still sees the node-pty kill it must not repeat. The in-process
provider's app-quit sweep keeps the blocking variant. The ps dialect
and table readers move to posix-pty-process-table.ts.

Refs #34410
Refs #33870

* fix(daemon): address review on async PTY process-table reads

Guard only the root-pid fallback on exit so proven groups are still swept
after dispose, wait out at most one ps dialect probe, move the sync-launch
baseline entry with the table reader, and keep tests cross-platform.

Refs #34410
Refs #33870

* fix(daemon): keep the PTY open until an in-flight force-kill sweep finishes

A sweep whose root exits mid-read must not signal: the freed TTY may
belong to another terminal. Dispose therefore defers destroy() until the
sweep settles, so its hangup cannot end the root first.

Refs #34410
Refs #33870

* test(daemon): build the destroy-tracking PTY mock without a type assertion

Refs #34410
Refs #33870

* fix(pty): make the local graceful-deadline force-kill async and stub ps in tests

The in-process provider's graceful-deadline timer now uses the async
group sweep; the app-quit sweep and the root-signal path stay
synchronous. Tests that drove force-kill through fixture pids now stub
the async process-table read instead of querying the host, and a new
test pins that waiters do not chain behind a failed dialect probe.
Session ids are only read for rows on the root's terminal.

Refs #34410
Refs #33870

* fix(pty): hold PTY teardown until an in-flight force-kill sweep settles

Orphan cleanup could destroy a local PTY while the graceful-deadline
sweep was still reading the process table, neutralizing its fallback
kill and hanging up the root first. Both the daemon handle and the
local provider now defer teardown through PendingPtyGroupSweep.

Concurrent force-kill callers in the daemon now await the shared
in-flight request, so a joiner's physical-exit budget starts after the
signals are sent and it sees their rejection.

Refs #34410
Refs #33870

* fix(pty): let disposal and orphan cleanup join the in-flight force-kill sweep

Session disposal now joins the controller's shared force-kill request
instead of starting a second sweep. Local orphan cleanup decides its
escalation only after a pending graceful-deadline sweep settles, so a
refused sweep is escalated before the handle is released rather than
leaving the retry with a neutralized kill.

Refs #34410
Refs #33870

* fix(daemon): SIGKILL live PTY groups before a shutdown deadline exit

The old synchronous sweep sent its signals before the shutdown deadline
timer could run; the async sweep can still be reading the process table
when it fires. The deadline now SIGKILLs each live session's process
group synchronously before exiting (node-pty roots lead their own
group), falling back to the root pid when the group kill is refused.
POSIX only.

Refs #34410
Refs #33870
2026-10-27 18:36:30 -07:00
Neil 5784e146de feat(wsl): internal distro allowlist that keeps WSL execs out of other distros on shared test boxes (#34539)
* feat(wsl): internal distro allowlist for WSL execs on shared test boxes

ORCA_INTERNAL_WSL_DISTRO_ALLOWLIST (comma-separated distro names) keeps every
wsl.exe call out of distros not on the list: process-host's resolveSpawn,
the main and daemon node-pty spawns, and the two codex-accounts calls that
bypass process-host. Pure list/status/version calls stay allowed; any other
form must name an allowlisted distro, and --shutdown or a default-distro exec
is refused. A malformed value refuses every distro and logs once. Unset, the
check reads the variable and returns, so behavior is identical to today.

* fix(wsl): cover every direct wsl.exe exec with the internal distro allowlist

- Restore #34525's worktree-create payload and captured-host test, reverted by
  accident in the first commit.
- Check the allowlist in execFileCapture, the choke point for WSL Git, gh and
  worker-thread captures, and before the WSL git-read probe, the Codex hook
  path canonicalization and the CLI login spawn. The WSL watcher now spawns
  through process-host.
- Stop scanning wsl.exe argv at the first positional argument, which wsl.exe
  runs in the default distro.
- check-process-host-imports fails on any direct wsl.exe exec outside
  process-host that does not call assertWslDistroAllowed.

* fix(wsl): check the distro allowlist where resolveCommand produces the wsl.exe form

Every Git, gh and glab path into a distro resolves through resolveCommand, so
gitExecFileSync, gitSpawn, git stream reads and wslAwareSpawn are covered too;
the execFileCapture check is now redundant and removed. The import check also
treats buildWslExecArgs as a wsl.exe exec and accepts resolveCommand as covered.
2026-10-27 16:19:13 -07:00
Neil ad5f787ad0 chore(process-host): drop the WSL hook relay launcher from the direct-import baseline (#34078) 2026-10-26 15:24:14 -07:00
OrcaWinandOrcaWin f34c13ca1e test: remove low-value main-process suites and browser probes (#33889)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-10-26 12:48:28 -07:00
Neil d4cab5c567 refactor: group local filesystem guards and search owners (#33624)
* refactor: group local filesystem guards and search owners

* refactor: use process-host handle contract for listing cleanup

* test: preserve concrete WSL module exports in search fixtures

* test: remove unnecessary WSL owner mocks from search fixtures
2026-10-25 13:40:31 -07:00
Neil d81958f0e8 refactor: group local and WSL filesystem watchers (#33407)
* refactor: group local and WSL filesystem watchers

* test: isolate startup ordering from watcher initialization

* test: collect the startup preflight subject once
2026-10-25 00:13:36 -07:00
Neil 95df4c74ed refactor: group Parcel watcher process ownership (#33401) 2026-10-24 22:14:52 -07:00
Neil b5ee026111 refactor: group Parcel watcher child process ownership (#33079) 2026-10-24 00:47:31 -07:00
NeilandBingxin Xu cd2eeebe3e perf(ssh): opt-in proxy connection reuse for system SSH on Windows (#33030)
* perf(ssh): reuse one proxy connection for system SSH on Windows

Windows OpenSSH cannot multiplex (ControlMaster fails with "getsockname
failed: Not a socket"), so every system-SSH exec for a ProxyCommand or
ProxyJump target re-ran the proxy handshake. Hold one `ssh -N -L` through
the real proxy per session and dial later execs over 127.0.0.1 with
HostName/Port/ProxyCommand=none/ProxyJump=none/HostKeyAlias overrides.

The route starts after connect, is proven with a login before use, falls
back to per-command proxying on any failure, honors
systemSshConnectionReuse: false, and is torn down with its proxy child.

Rebased from #31731 (commits b1cdd46, 3835c2d, 6268f2f) onto current main,
including the pending-route cancellation fix from b301b6a.

Co-authored-by: Neil <neil@stably.ai>

* refactor(ssh): stop the loopback forward through process-host's tree kill

Reuse signalProcessTree instead of a hand-rolled taskkill-then-kill
fallback: it already joins an in-flight tree kill, refuses a reaped PID,
and falls back to killing the root, so the route keeps no parallel copy
of that logic.

* fix(ssh): harden the Windows loopback route against review findings

- Manual targets: take the forward's sshd port and the HostKeyAlias from
  the stored endpoint; `ssh -G` ran on the label and named another host.
- Hold the forward with ServerAlive probes so a silently dead proxy ends
  it and commands fall back instead of stalling on 127.0.0.1.
- Never route forwards: their argv must stay stable for orphan recovery,
  and the route's own forward must take the real proxy.
- Upload sessions read the route per file, so an exited route's port is
  not dialled after it closes.
- Skip the route for security-key identities, whose forward and probe
  logins would each ask for an unprompted touch.

* fix(ssh): keep the loopback route for manual targets no Host block claims

The unclaimed-alias endpoint rewrite only applies to config-backed targets,
so manual targets with a configHost no block names stay eligible.

* fix(ssh): keep the held loopback forward non-interactive

The forward is an optional background login; with BatchMode=no a password or
passphrase login raised an extra prompt right after connect, which the
non-interactive probe then discarded anyway.

* fix(ssh): admit the loopback route only when it changes nothing but the route

Overriding HostName to 127.0.0.1 changes which ssh_config blocks match, so a
`Match host <real host>` block could silently switch the remote User (or
identities, certificates, known-hosts files) for every routed command. Run
`ssh -G` with and without the route options and refuse the route on any
difference outside HostName/Port/HostKeyAlias/ProxyCommand/ProxyJump.

Also ask the agent with `ssh-add -L` and refuse the route when it may hold a
security key, so the background forward and probe never request a touch.
Both checks run through process-host and fail closed.

* feat(ssh): make Windows proxy connection reuse an opt-in per-target setting

The loopback route rewrites HostName, which ssh_config can key settings on,
so it no longer turns on by itself: a new optional systemSshProxyConnectionReuse
field (default off, Windows only) enables it from the SSH host's Advanced
settings. Old clients and servers ignore or omit the field.

Admission stays as the safety net and now:
- takes the forward's sshd port and HostKeyAlias from the direct `ssh -G`,
  so a manual host's own Host block cannot send the route to another sshd;
- refuses any non-routing value with a `%` token, which `ssh -G` can leave
  unexpanded while a routed login expands `%h` to 127.0.0.1;
- checks the security-key and agent identities of every ProxyJump hop;
- asserts the routed config really dials 127.0.0.1 with no proxy left.

* fix(ssh): offer proxy reuse only for hosts this desktop dials, refuse ${VAR} paths

- The switch now appears only for this desktop's own SSH hosts. A paired
  server's add-host dialog never shows or sends it, so an older server's
  strict schema is never handed the new field, and a client's platform no
  longer decides a server's setting.
- Admission also refuses `${VAR}` references, which `ssh -G` prints
  unexpanded while a login expands them.
- Advanced no longer auto-opens just for the opt-in.

---------

Co-authored-by: Bingxin Xu <xubingxin.xbx@gmail.com>
2026-10-23 21:22:08 -07:00
Jinwoo Hong d3dbd9683f Update Orca's skills with Orca's own installer instead of npx skills (#31688)
* Add Orca's own bundled-skill planner and installer (not wired in yet)

One planner decides what installing Orca's bundled skills does to each
target folder; execute re-plans under per-path locks and runs only a plan
whose digest still matches; a shared summary derives badge, wording and
exit code. A copy counts as Orca's untouched one only when its SKILL.md
bytes match a version some build shipped (resources/skills/released-digests.json,
generated from git history and checked add-only).

Nothing calls this yet; the npx path is unchanged.

* Report left-alone folders as warnings and keep backups in the home

A leave is no longer a failure: the outcome line says what happened plus
"N left alone: <reason>", and the exit code is 1 only for a failed or
changed op. Backups for every scope go to <home>/.agents/.orca-skill-backups
on the executing host, so a project Replace never writes into the repo;
moving there falls back to copy, verify, then remove across filesystems.

* Update Orca's skills with Orca's own installer instead of npx skills

The app's Update skills dialog, badge, nudge and status pills now read the
bundled-skill plan and summary, and Update executes the reviewed plan's
digest. `orca skills install|update` runs the same planner in-process with
--dry-run, --force, --json and per-summary exit codes. Removes both npx
spawn sites, the freshness inventory modules and the npx round-trip CI.

* Update only where the skill already is; one provider-root rule for app and CLI

installedOnly now drops every location that holds nothing (keeping the shared slot
only when an existing agent link points at it), so Update never adds a skill for
another agent. The app and CLI resolve the Claude skills folder through the same
rule in localBundledSkillInput; a test asserts both plan identically.

* Compact Update skills dialog: one row per skill, locations named by agent

summarizeBundledSkillPlan now returns each skill's locations, labelled by
agent (or "Shared folder") with a ~-collapsed path; the dialog shows one row
per skill that needs something, collapses up-to-date skills into one line, and
lists every location behind "Show all locations". The result reads "Updated N
skills · M left alone". Under installedOnly a Replace with no shared copy
replaces in place instead of adding one.

* fix(skills): keep location paths readable by putting the state under the path

* fix(skills): hold the last plan in state, not a ref mutated during render

Split the update dialog's rows, headline and footer into components, move the Update label
next to the other plan copy, and use z.strictObject for the request schemas.

* Escape non-ASCII characters in bundled-skill regex and wording

The orcad bundle rejects a literal non-ASCII character in a regex.

* fix(cli): escape non-ASCII apostrophes in the skills install help text

* test(orcad): check the bundled CLI plans a skill install without npx

Drop the deleted roundtrip workflow's group from the closed-PR cancellation fixture.

* Make bundled-skill test fixtures link folders the way the product does on Windows

Directory links in fixtures are junctions on Windows, which need no
Developer Mode; the one file-link fixture is skipped there.

* Resolve legacy skill agent keys canonically

* Reject stale skill replacements when auxiliary content changes

* Keep ordinary bundled skill classification shallow

* Leave unreadable skill replacements alone without aborting plans

* Pass force-all through shallow planning and describe unreadable copies

* Clarify skill update statuses and folder details

* Keep selected skill replacements in place with one status

* Report completed skill replacements with their backup confirmation

* Clarify available skill updates and skipped copies

* Keep skill update results in preview order

* Show each skill skip reason once

* Count skipped skills instead of skipped folders

* Make installer input and link tests work on Windows

* Bound bundled skill classification reads and parse names linearly

* Keep workspace skill roots inside the resolved workspace

* Compare workspace skill roots with native path semantics

* Restrict bundled skill backups to their owner on POSIX

* Keep recovered skills needing attention until they are updated

* Count completed skill updates in the status tooltip

* Sync the runtime localization catalog
2026-10-22 22:35:14 -04:00
Brennan Benson 2db7d5aac4 Refuse chat resume until terminal writers have exited (#29242)
* fix(native-chat): identify the owner when resume create is refused

* test(native-chat): consolidate transcript refusal cases

* test(native-chat): keep ownership coverage at the launch boundary

* test(native-chat): update adoption fixture for early ownership admission

* test(native-chat): align resume admission assertions with host ownership

* test(native-chat): keep RPC coverage focused on ownership proof

* test(native-chat): assert blank request presence before optional source

* fix(native-chat): admit the actual CLI source during chat creation

* fix(native-chat): satisfy source admission lint checks

* test(native-chat): provide the waiting hook prompt

* test(native-chat): isolate source admission fixtures from host homes

* test(native-chat): canonicalize adoption fixture directories

* test(native-chat): preserve early ownership admission in source stack

* Resume CLI conversations directly in native chat

* fix(native-chat): preserve resume refusal behavior and honest transcript feedback

* fix(native-chat): retain import read evidence in typed result

* Use the header test fixture record type

* Consolidate launch owner imports and correct feedback assertion

* Keep existing resume translation in runtime catalog

* fix(native-chat): preserve admitted source and prepared execution home

* fix(chat): retain source evidence for dispatched adoption

* fix(chat): keep source capture within module limits

* Keep deferred chat publication in the shared launch owner

* test(chat): allow host evidence in create routing

* Satisfy launch publication checks and scope slot cleanup

* Preserve canonical tab membership on projected reorders

* Use recorded launch ownership for surface recovery

* Reserve deferred chat order through existing group actions

* Pass the active layout owner into launch cancellation

* Stop terminal writer before adopting conversation in chat

* Guard chat adoption completion and satisfy static checks

* Switch completed terminal turns without confirmation and test handoff

* Recheck terminal writers on reserved adoption retry

* Split session entry delegates and fix linted test types

* Keep chat opening send below create module line limit

* Require explicit terminal stop consent before chat adoption

* Use direct process-table test mock

* Fix chat adoption types and runtime localization

* Localize terminal takeover refusal guidance

* Keep CLI stop admission scoped to active source panes

* Preserve ordinary history resume through attach

* fix(native-chat): admit an unsaved transcript under a linked account folder

A missing transcript stayed unresolved while its account root resolved through
symlinks, so containment failed and a not-yet-saved conversation showed a load
error instead of the unsaved notice.

* Update adoption replay test for at-rest creation

* refactor(native-chat): move the newest-item import window into its own module

Keeps journal-legacy-import.ts under the line limit after the source
identity observer landed on top of the streaming import.

* fix(native-chat): restore the refusal error import lost in the main merge

* Remove duplicate account resolver and condense terminal admission

* Refuse conversation adoption until terminal writers exit

* Keep conversation inspection within lint limits

* Preserve conversation owner in launch feedback

* Expect unavailable writer scan to refuse fresh adoption

* Keep terminal conversation evidence tied to its execution

* Correct writer continuity test refusal and event shapes

* Prove original terminal writers exited before chat adoption

* Scope saved conversation holds to their execution host

* Keep child-inspection fixture verdict literal

* Recheck awake saved conversation holds against their owned pane

* fix: inspect owned Windows jobs independently of foreground naming

* test: use braces in Windows inspection fixture

* test: complete ordinary Windows terminal inspection fixture

* test: explain partial terminal inspection fixture

* fix: inspect the terminal a conversation is switched out of

* test: separate OpenCode writer observation coverage

* fix: tie terminal conversation links to the PTY execution

The conversations a terminal reported now live on its PTY record for that
incarnation, so moving the pane or running many other terminals can no longer
drop them, and a link whose terminal can't be inspected refuses.

- Each link keeps the agent process its hooks last proved; a live or
  suspended process refuses even when the terminal's tree is shell-only
  (detached tmux, nohup).
- tmux-published events are never pane-owned, and an unmarked binding copy no
  longer revokes an earlier pane-owned proof.
- A restored row or saved checkpoint is censused against the pane's current
  local PTY instead of requiring this runtime's observed identity.
- The refusal now says Orca can't confirm the agent has exited and asks the
  user to stop what's running there or close the terminal.

* test: drop the stub-host admission calls from OpenCode observation coverage

* test: census suspended children through the real terminal host

* test: drop writer checks the real-terminal flow already covers

* fix: let prompt workers outside the terminal's session pass the exit check

- The POSIX shell-alone proof skips a session leader below the shell and
  everything under it: a prompt's zpty worker has left the terminal, while
  suspended, background, nohup and disowned jobs still refuse.
- Windows steady-state polls read the cached process table; every read that
  decides adoption or close stays fresh.
- The pane header and tab menu point back to Session History, since their
  action is gone once the agent exits.

* fix(terminal): count only Windows job members still linked to the shell

A pane's job keeps every process it ever started, including daemons whose
parent exited (an rc-started ssh-agent), so such panes never read as an idle
shell and the chat switch refused forever. Walk descendants of the shell along
live parent links within the job, for every shell; the Git Bash launcher keeps
its bash-chain allowance and a child older than its parent still refuses.

* Revert "fix(terminal): count only Windows job members still linked to the shell"

MSYS exec and fork leave intermediate Win32 parents exited, so a live Claude
started through a script (and 'sleep 60 &') became unreachable and the chat
switch admitted while it ran. Back to the job census until a console-based
design is measured.

* fix(terminal): set aside detached daemons on a Windows shell-alone decision

A pane's job keeps every process ever started in it, so an rc-started
ssh-agent made a Git Bash pane read as busy forever and the chat switch
refused. MSYS fork/exec leaves live work under exited Win32 parents, so parent
links alone can't tell the daemon apart; it is the one that also left the
console. On a scanChildProcesses decision read (switch admission, run-terminal
close, close guards), a member neither linked to the shell nor attached to its
console no longer counts. Polls and shell confirmation keep the job census.
The console-list helper now forks through the process-host chokepoint.

* refactor(daemon): name the process inspection options once

Keeps the daemon router and terminal host within their line budget now that
the options carry scanChildProcesses.

* fix(terminal): set aside detached daemons only for adoption, never for close

Closing a Windows pane kills its whole job, so a close check that ignored a
detached process (gitk &, cmd /c start ...) would kill it without asking.
Adoption asks whether anything is still in the terminal; close asks whether
closing destroys something. Only the adoption read now sends ignoreDetached;
close guards and the run-terminal sweep keep the job census. The win32 test
also stops the ssh-agent it starts.

* fix(native-chat): check the terminal being switched out of before opening its chat

When the conversation already had a native chat, the switch hit the
ownership conflict first and the header opened that chat, even while the
terminal's agent (a later claude --resume) was still running. The switched-out
terminal is now checked before the conflict, so a live one refuses as usual
and one at rest still opens the owning chat.

* test(terminal): assert close reads never call an ssh-agent pane idle

A plain read can answer unverifiable while the eval's short-lived processes
exit; either way it must not read as no-children.

* fix(native-chat): check terminal writers before naming an owning chat for any switch

Session History's Resume in New Native Chat carries no source pane, so with an
earlier chat still holding the conversation it skipped the terminal check and
the owning chat's first send would start a second writer beside a live
terminal Claude. Every adoption request now runs the writer check before the
owner conflict, as the no-owner path already does.

* test(reliability): run the chat-switch writer oracles in the provider-ownership gate
2026-10-19 01:06:36 -07:00
NeilandClaude 139bac88b6 refactor(ssh): move the ssh-connection and ssh-target cluster into ssh/connection/ and ssh/targets/ (#30479)
* refactor(ssh): move the ssh-connection and ssh-target cluster into ssh/connection/ and ssh/targets/

Pure git mv plus import and path-literal rewrites. ssh-connection-utils stays put; the next PR splits it.

Co-Authored-By: Claude <noreply@anthropic.com>

* test(e2e): keep the release-checkout import at the v1.4.218 path

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>
2026-10-18 19:23:21 -07:00
Neil 658953dc23 refactor(ssh): group system-ssh transport, ssh config and remote exec into ssh/{system-transport,config,remote-exec}/ (#30355)
git mv plus import, vi.mock and path-literal rewrites; one e2e route regex follows the moved
dynamic-forward process.
2026-10-18 07:26:10 -07:00
Neil 932dc1be29 chore(runtime): delete the never-served TLS branch of the LAN WebSocket server (#30220)
No released build ever passed a certificate to WebSocketTransport or set the pairing TLS
fingerprint; only ws-transport's own tests used tls-certificate.ts. The tests now run over ws://.
2026-10-17 23:44:42 -07:00
Neil d4aa9dcadc revert: restore pinned npx skills installer (#30207)
Reverts #30185 so the separate installer project can proceed. The restored shared command already pins skills@1.7.2 for CLI, desktop updates, and setup commands.
2026-10-17 23:24:02 -07:00
Neil 4baacca5c0 fix(startup): shell PATH probe falls back and retries when the shell cannot launch (#30204)
* fix(startup): shell PATH probe falls back and retries when the shell cannot launch

* chore: drop hydrate-shell-path from the process-host direct-import baseline
2026-10-17 22:54:54 -07:00
Neil ffb818bf61 Ship the locked skills installer instead of executing npx (#30185)
* security: ship a locked skills installer instead of running npx

* chore: remove retired skills process import exception

* test: include the locked installer in packaged layout fixtures

* test: reuse the real installer builder for package fixtures
2026-10-17 22:51:14 -07:00
Neil 7f8faae800 feat(space): open disk usage in a separate window (#29513)
* feat(space): open disk usage in a separate window

* perf(space): move local disk scan off Electron main thread

* fix(space): satisfy full PR quality gates
2026-10-16 01:25:19 -07:00
Žan MlakarandOrca Worker 88bd5db42e fix(windows): copy a file to the clipboard without a visible PowerShell window (#28781)
* Run clipboard file copy and TLS openssl through process-host

* refactor(clipboard): inject runProcess, pass the Windows path as data, add an e2e

- clipboard-file-copy builds the process spec itself, so the stdio policy sits
  next to the wl-copy/xclip knowledge it exists for; the IPC handler passes
  process-host's runProcess straight through and keeps no wrapper.
- A helper's exit is read as data (clean exit or not) instead of being turned
  into a throw and caught again.
- Windows runs a constant Set-Clipboard script and reads the path from the
  environment, so no path is ever parsed as PowerShell.
- Drop the dead tls-certificate rewrite; nothing in the app imports it.
- e2e: right-click Copy on Windows, then a real shell paste of the file.

* ci(e2e): run the Windows explorer file-copy spec on the Windows runner

Route it from the clipboard sources, keep it out of the Linux lanes, and run it
in the existing Windows e2e job beside the other Windows-only specs.

* refactor(clipboard): tighten the helper seam, tests and e2e after review

- runClipboardHelper takes runProcess and a plain ProcessSpec.
- One handler test pins the process-host wiring on every host; spec details
  stay in the unit tests, which now also pin the inherited environment.
- e2e: native realpath so an 8.3 runner temp path still matches, paste via
  process-host, and no toast check that could never fail.
- Route the renderer caller to the e2e too.

---------

Co-authored-by: Orca Worker <orca-worker@localhost>
2026-10-15 14:22:27 -07:00
OrcaWinandOrca Worker 4559c250b3 fix(windows): restrict secure files in-process instead of spawning icacls (#28135)
* fix(windows): restrict secure files in-process instead of spawning icacls

Every secure write on Windows spawned icacls.exe five times on the calling
thread (the Electron main thread for most stores): verify, /reset,
/inheritance:r /grant:r, verify, verify. Process creation under endpoint
security costs ~0.6 s each, so each write froze the UI for seconds (#26027).
The /reset + /inheritance:r pair is also what behavioural EDR scores as
ransomware tradecraft (#25583).

Add native/windows-security (@orca/windows-security): a small N-API addon
over advapi32 that reads and writes DACLs in-process. It exposes the token
user SID, a DACL read as SDDL, a DACL write (sync and on the thread pool),
a create-with-DACL file write, and an additive grant. Policy and
verification stay in TypeScript.

- writeSecureFile creates each temp file already carrying the protected
  user/SYSTEM/Administrators DACL (CreateFileW + SECURITY_ATTRIBUTES), reads
  it back once, and renames it. No ACL is modified on the write path and no
  process starts. Before: 5-7 spawns, 94-138 ms per write. After: 0 spawns,
  ~0.5 ms.
- A secure directory is verified once per process. It is rewritten only
  when verification fails, on the thread pool, because the write walks
  every descendant (~2 s on a 29k-entry userData).
- The verifier follows OpenSSH's rule: only the user, SYSTEM and
  Administrators may be granted anything, and the path must be protected
  (or, for a file, inherit from a secure parent). DACLs written by the
  icacls-era hardening pass unchanged, so there is no migration.
- If the addon cannot load, the write still happens and reports
  unrestricted, logged once. There is no icacls or PowerShell fallback.
- grantDirAcl, the startup userData grant, and the install-dir probe and
  package repair use the same addon, so whoami.exe and icacls.exe are no
  longer spawned anywhere at runtime.
- orcad's Windows slots ship the addon beside their bundles (built per
  arch, PE-machine checked). Its data-root privacy check fails closed, so
  without the addon a Windows orcad would not start.

Fixes #26027
Fixes #25583

* refactor(secure-file): harden existing Windows files through the sync lane

The read path called restrictWindowsPathSync directly and recorded success
itself; the existing sync lane already does both. One call site fewer.

* refactor(windows-security): one load, one rule, no fallthrough (review #28135)

- The descriptor parser is anchored, so a conditional ACE (its body holds
  parentheses) or anything else it cannot read fails closed instead of
  verifying the ACEs before it.
- findSecureDaclViolation is one flat rule: protected, allow rules only for
  the user, SYSTEM and Administrators (OpenSSH's ACE rule), and full
  control for the user. The inherits-from-a-secure-parent exception is gone:
  it accepted exactly what loosening the parent reaches.
- The addon and the user SID load once, behind one wrapper that adds the
  extended-length prefix to every path and runs async DACL writes to one
  path in order, so a startup grant cannot land after a re-lock and restore
  the broader DACL.
- With the addon loaded, a failed born-restricted create throws with the
  matching Node error code instead of falling back to a plain write.
- A grant leaves a NULL DACL untouched instead of turning 'everyone' into
  'only this SID'. WriteFile runs once and fails on a short write.
- The artifact create-intent store writes through writeSecureFile and drops
  its own temp writer, directory LRU and synchronous directory re-lock.
- Test seams the global fake makes redundant are gone.

* fix(packaging): fail a Windows build whose workspace addon binary is missing or foreign-arch

The @orca/windows-* workspace links exist on every host, so their presence
proved nothing: a missing or wrong-architecture .node only surfaced at
runtime, as unrestricted secure files or an orcad that refuses to start.
beforePack now reads each addon's PE machine against the target arch.

* refactor(windows-security): prune what the native ACL left dead (review #28135)

- Sync-spawn ratchet: hardenSecurePath and the sync-flag exemption it alone
  used are gone, and the advice and baseline header no longer say secure
  writers run icacls.
- The install-dir probe reads ACEs through parseSddlDacl, so an ACE it
  cannot parse fails the whole reading closed instead of being skipped; AC
  is a resolved alias there now.
- The packaging check reads the addon binary through node_modules, the copy
  that ships.
- No winCode-to-errno table: nothing reads .code off a failed secure write.
- The load memo is not mutated; secure-path keeps its own log-once flag.
- grantInstallDirPackageAcl is a plain function, and secure-file restricts
  a Windows directory and a POSIX path directly instead of through one
  two-platform helper.

* fix(windows-security): resolve LA in real-ACL tests; tear a file whose close fails

- The GitHub Windows runner is the built-in Administrator (RID 500), whose
  SID reads back as LA. Production already resolves it through
  localDomainSidOf; the win32 test helpers parsed without it. One shared
  readDaclOf resolves aliases for every real-ACL suite, and the verifier's
  unit test covers an LA-aliased directory and LA as a stranger.
- createFileWithDacl treats a failed CloseHandle like a failed write (lost
  data on a network volume) and deletes the torn file.
- Drop a duplicate test reset and a typo.

* fix(windows-security): load only from inside the app; read DACLs as structure (review #28135)

- Loader: orcad's staged binary first, then the workspace package only if
  its real path is inside the app (process.resourcesPath, or this
  checkout). A bare specifier otherwise walks every parent up to
  C:\node_modules, which any local user may create, and orcad holds
  credentials.
- The addon reads a DACL as { protected, aces: [{ type, flags, mask, sid }] }
  via GetAce, with a SID only for plain allow/deny ACEs and null for a NULL
  DACL. findSecureDaclViolation and the install-dir probe check that
  structure, so conditional, callback and object ACEs fail closed. The SDDL
  parser, its alias table, LA/localDomainSid resolution and getDaclSddl are
  gone; SDDL is only the write format.
- Install-dir tree pass: files flagless as before, but a directory only when
  its own DACL is protected (inheritably). Re-granting inheriting
  directories re-propagated each subtree (3.2 s for 1,093 directories). The
  root is no longer granted, or counted, twice. The two passes are plain
  functions behind a { root, tree } seam.
- One namespacing (path.win32.toNamespacedPath) at the addon boundary, also
  the per-path ordering key; the comment states the real sync-grant
  invariant.
- restrictWindowsDirectory(dir): Promise<boolean> replaces the callback
  lane; the reporter is guarded once, so the 'settle' stage is gone.
- Sync-spawn ratchet: the secure writers no longer spawn or block, so they
  leave the gate (109 -> 61 uses).
- Errors name the Win32 call and code only; a failed
  GetSecurityDescriptorControl fails the write instead of defaulting to
  unprotected.

* test(secure-file): assert the durable POSIX write's outcome, not the old chmod order

writeSecureFile now chmods the temp file to 0o600 before fsyncing it, so
the fsync also covers the mode change and the file is owner-writable when
opened. Assert what matters under umask 0o200: the write succeeds, the
temp file is fsynced read-only, and the published file is exactly 0o600.

* fix(windows-security): namespace paths with the host's toNamespacedPath

path.win32.toNamespacedPath rewrote POSIX temp paths (/tmp/x -> \tmp\x) in the
Linux runs of the Windows-logic unit tests, so the fake addon wrote to the wrong
place. The addon only runs on Windows, where the host function is the win32 one;
elsewhere it is the identity. The long-path assertion runs on Windows only.

* test(windows-security): assert the install-dir long-path prefix on Windows only

The prefix comes from the host's toNamespacedPath, which is the identity off
Windows; the Linux run of this suite expected the win32 escape.

* fix(windows-user-data-acl): grant the root even when its children cannot be listed (review #28135)

A readdir failure on userData aborted the whole grant, so the root grant
that could repair a DACL refusing the listing was never attempted. The
listing failure is now recorded, the root is still granted, and the
marker stays withheld so the next launch retries.

* test(windows-security): wait for relock propagation, accept Entra ID SIDs (review #28135)

- The large-relock test waited only for the parent to turn protected,
  which happens before propagation reaches the children, then read a child.
  It now waits on the parent and both ends of the child range.
- The token-user test rejected S-1-12-1 (Entra ID) accounts, failing the
  suite on Entra-joined dev machines.

---------

Co-authored-by: Orca Worker <orca-worker@localhost>
2026-10-13 01:56:55 -07:00
OrcaWinandOrca Worker fb5d016887 fix(wsl): start every host wsl.exe call through process-host so no console window flashes (#27628)
* fix(wsl): start host-side wsl.exe probes through process-host so they never open a console

wsl --list, wsl --status, the WSL home and directory probes, and WSL UNC
deletes were started with node:child_process directly. The synchronous ones
omitted windowsHide, so each spawn from the GUI-subsystem main process
flashed a conhost window that could take keyboard focus. Route them through
@orca/process-host, which pins windowsHide, and keep the execFile-shaped exit
errors the availability cache classifies.

Refs #27390

* build(tc): list wsl-exe-process.ts in the web typecheck project

wsl-availability.ts is in the explicit web file list and now imports it.

* refactor(wsl): drop the unused sync availability probe; classify ProcessResult directly

No production code calls the sync isWslAvailable (the IPC and RPC paths
already assert they never do), so delete it and its sync guest-probe twin
instead of porting them. The async probe now reads the --status
ProcessResult itself: an exit code is definitive, a timeout or signal kill
is retryable, ENOENT on spawn is definitive. The WSL2-missing-kernel exit
codes live in one named set. Availability tests now drive the async probe.

* refactor(wsl): shrink wsl-exe-process to a spec builder and stdout reader

With the availability cache classifying ProcessResult itself, nothing reads
the execFile-shaped code/status error any more. wslExeFailure returns a
plain reason string; wslExeStdout throws it; the UNC delete folds it into
its own error. The home probe builds its argv with buildWslExecArgs.

* test(wsl): mock process-host in the empty running-list case from main

* docs(wsl): the availability probe no longer blocks; drop the stale word

* build(ratchet): prune the wsl.exe sync launches this branch removed

---------

Co-authored-by: Orca Worker <orca-worker@localhost>
2026-10-12 14:51:32 -07:00
OrcaWinandm4air 93e0b80be7 refactor(cli): serve children spawn through process-host (#27760)
Co-authored-by: m4air <m4air@Mac.localdomain>
2026-10-12 03:23:49 -07:00
OrcaWinandm4air 2b4b60a87e refactor(relay): delete the SSH relay server, its build assets and tests (#27215)
* wip: delete the SSH relay server and its build, assets and tests

* wip: drop main/relay plugin conformance; WSL reader lookup test

* fix: drop the emptied relay bench; prune baselines

* test: keep the native halves of tests that also covered the relay

* chore(ci): drop shard timings for the deleted relay server tests

* fix: keep the directory listing budget that darwin terminal names now import

---------

Co-authored-by: m4air <m4air@Mac.localdomain>
2026-10-10 22:36:42 -07:00
ff27338299 refactor(ssh): connect fails for hosts orcad can't serve; remove the relay client path (#27214)
* wip: connect fails for hosts the managed server can't serve; remove relay session, deploy and relay UI

* wip: delete relay-only client modules left unreachable, prune their tests and gates

* wip: format

* wip: exhaustive unserved message

* wip: drop relay attestation test for the deleted ownership proof

* fix: prune stale baselines; type the test transport stub

* fix(ssh): census proves leftover relay leases; name the remedy for older terminals; translate every connect-failure surface

* test(ssh): leftover leases defer to the host census

* test(e2e): remove relay-path SSH e2e specs and their CI lanes

* test(e2e): drop the auto-convert case that built its relay era through a live relay

* test(e2e): mark relay-era browser and catalog journeys fixme until they seed a profile

* test(e2e): seed relay-era browser and catalog journeys from a saved profile

The three journeys built their old-version state through a connect without an orcad template, which now refuses the host. They now start from a seeded relay-era profile, so the first connect converts the host as an upgrading user's would.

* chore(ci): drop shard timings for the deleted relay tests

* test(e2e): carry a relay-era login cookie across the browser conversion; own every launch

* test(ci): drop the deleted SSH layout oracle from its route

* test(e2e): open the retained browser after conversion, retrying a pre-connect card

* test(e2e): bind the relay-era browser partition before seeding its cookie

A relay-era build always recorded the partition binding before a page stored
data in it. Seeding the cookie into an unbound partition made the binding store
refuse it (browser_route_partition_binding_store_invalid), so the retained tab
showed the generic routing card and Retry could never clear it.

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: m4air <m4air@Mac.localdomain>
Co-authored-by: Claude <noreply@anthropic.com>
2026-10-10 22:10:33 -07:00
OrcaWinandm4air b4d8e370da fix(orcad): install agent status hooks when a managed host starts (#27319)
* fix(orcad): install agent status hooks when a managed host starts

* fix(orcad): read login-profile Grok/Kiro homes before installing hooks

* refactor(agent-hooks): run the login-shell home probe through process-host

---------

Co-authored-by: m4air <m4air@Mac.localdomain>
2026-10-10 17:08:36 -07:00
OrcaWinandm4air 31b2b17d9b refactor(wsl): move the WSL guest bundle sources out of src/relay (#27174)
* refactor(wsl): move the WSL guest bundle sources out of src/relay

The WSL hook relay and WSL browser network relay are built from src/relay but are
not part of the SSH relay. Move them, and the 46 relay modules the hook relay
bundles, to src/wsl-guest so the SSH relay can be deleted on its own. Pure
renames; the built bundles are byte-identical.

Adds a boundary test so nothing outside src/relay imports it except relay tests.

* fix(ci): exempt renamed files' existing lint debt in the changed-lines gate

* test(relay): keep the cast suppression on the wrapped line

* test(relay): only parse files that mention relay in the import boundary check

---------

Co-authored-by: m4air <m4air@Mac.localdomain>
2026-10-10 02:46:27 -07:00
Neil 1912c90141 Extract subprocess handling into @orca/process-host (#26963)
Extract subprocess execution and process contracts into @orca/process-host with explicit public entry points, independent checks, and one Node 18 compatible CommonJS runtime. Migrate application consumers and integrate package builds, dependency selection, import enforcement, and runtime resource staging.

Preserve daemon startup ownership and profile isolation, and qualify development rebuild recovery and concurrent output publication. Remove unnecessary aliases, test forwarding exports, and a duplicate output scan.

Validation: full typechecks, changed-code quality, package and artifact checks, native Windows builds/types plus 306 runtime tests, and independent native reviews.

Matched full E2E controls on original main and unchanged reviewed head both pass all 17 applicable tests (nine skips). Earlier intermittent E2E failures are documented in the PR; no assertions were weakened. The final combined-tree audit preserves both branches with no import-boundary violations.
2026-10-09 18:32:28 -07:00