mirror of
https://github.com/stablyai/orca.git
synced 2026-10-31 17:17:38 +00:00
d3fc80292be694f454717e18f82fdc19dbe353de
11
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Brennan Benson
|
094f589977 |
Floating-workspace chats of every native-chat agent keep links in the folder they started in (#26981)
* fix(native-chat): resolve every registered agent's floating chat to its pinned folder Drop the unused showStructuredAgentSessionChoice and say what the built-in agent check means. * fix(native-chat): launch floating OpenCode, Grok, OMP and Pi chats in their pinned folder The renderer now resolves every floating chat's links, images and skills against the folder the chat was pinned to. The ACP (OpenCode/Grok/OMP) and Pi launchers still started the agent in the current floating folder setting, so after the setting moved and the agent relaunched, the agent ran in one folder while links resolved against another (and Pi forked its conversation into the new folder). Both launchers now use resolveAgentSessionLaunchDirectory, the helper Claude and Codex already use: launch in the pin, refuse with the existing "folder this chat ran in no longer exists" message when it is gone, and pin an unpinned floating record on first launch. Worktree and folder chats are unchanged. The tab-directory reader now calls the status bridge's own isStructuredTab instead of restating its predicate inline. |
||
|
Brennan Benson
|
573ff0bca3 |
Save a new chat's first message during create (#26332)
* fix(native-chat): the host owns the send queue; the window keeps no saved outbox The desktop kept each structured chat's unsent messages in localStorage and sent them in order, so one message whose fate was unknown froze every later send, Retry dropped it silently, failed sends could not be discarded, and an offline chat could send hours later. The host already records every message and owns the queue; the window now only sends. - One in-memory sender for composer, launch prompts and messages sent from outside the chat. One send in flight per chat; a transport failure resends the same id for up to 30 s; a refusal that proves nothing was recorded puts the text back in the composer with the reason; a send that went out and was never answered shows an in-doubt line with Send again and holds nothing up. - The host's "unknown" rows get the same in-doubt line and Send again, from the journal, in every window. - A message an older build left in localStorage is never sent: the host's conversation outline decides what goes back to the composer, and the copy is deleted once that is saved. * fix(native-chat): send through the structured chat RPC wrapper, with its timeouts The sender called the runtime RPC directly, skipping the per-method timeouts every other structured chat call gets. Only a remote host's request skips the compatibility check the sender already ran. * fix(native-chat): an unconfirmed send goes back to the composer, with no new row line The common pattern draws nothing extra on a message whose delivery is in doubt once its turn is over, and puts a failed send's text back in the composer with the reason. So a send nobody answered in time, or one a host answers in a way that proves nothing, goes back to the composer worded as unconfirmed, and a host "unknown" row shows nothing extra. Removes the in-doubt phase, Send again, and its strings. A host's made-up row for an id its journal lost now reads as unconfirmed, not as recorded, so that message comes back instead of vanishing. * test(native-chat): pin that nothing resends a send given back as unconfirmed * fix(native-chat): sends survive a tab close, never resend after a Stop, and keep remote images - A normal tab close lets sends on their way settle; what the host never took goes back to the conversation's draft. Only a cancelled launch or a worktree purge drops them. - After any Stop, a send already on its way is never sent again under its id: a doubtful answer, or a resend that was due, hands its text back worded as unconfirmed. - A returned image keeps the SSH connection it lives on; the connection never goes to the host. - An older build's saved message the host recorded and then rejected is left to the host's own row, never handed back. * fix(native-chat): a Stop or a tab close never resends a send already out, and a kept card is the card's A Stop that landed while a same-id resend was being readied (its timer fired, its request not out yet) let that resend go out after the Stop. The sender now tracks whether a request is awaiting its answer: a Stop hands back every send between attempts as unconfirmed, lets one whose request is out settle from its answer, and never issues a request after it. A normal tab close withdraws the same way instead of doing nothing, so nothing more goes out and nothing is dropped. A send the host rejected but kept as a card (keptAsQueuedMessageId) is the card's, from its reply or the journal, and never goes back to the composer. Adds the freeze tests: a send whose fate is unknown holds later sends no longer than its deadline, and one the host can neither confirm nor deny releases the next at once. * fix(native-chat): read a resend's turned-away call or reused id as unproven Ports the send-answer proof contract. A call the host turned away before running it (method_not_found, invalid_argument, unauthorized) proves only that this request wrote nothing, so it reads as never sent on a first attempt only; on a resend an earlier attempt may have landed, and it goes again under the same id. A resent id the host says was already used for other content (messageIdReused) proves nothing either, like an expired or conflicting id. Pins the rest of the contract: any row the host returns is its own, a thrown error is no answer whatever its code, and an older build's refused, held or outlived-Stop copy is handed back once and never sent. * refactor(native-chat): type the structured composer's send with its attachment type Keeps NativeChatStructuredSession.tsx within the file length limit. * refactor(native-chat): drop the kept-card guards the sender never needed A kept send's row is a rejection that was never a Stop's, so the sender already reads it as recorded, from its reply or the journal. The test that pins it stays; the two extra checks only covered a kept row that is also a withdrawal, which the host never writes. * test(native-chat): route launch tests' sends through the client wrapper the sender calls The sender sends through callStructuredAgentSession, but these launch tests replaced that module with a factory that answered nothing (and still named a probe that no longer exists), so every launch prompt resent until its 30 s deadline and the tests timed out. Each factory now hands sends to the runtime RPC mock the tests already answer, and expectations of a local send no longer ask for the remote-only compatibility option. * fix(native-chat): hand a message back without importing the composer's attachment hook The worktree purge reaches the launch prompt, which hands text back, and the attachment hook's imports reach the store. A test that builds the real store behind a mocked one then waited on itself and hung. Handing back now writes images to the draft store directly, as the hook's helper did, and a test pins that each returned image keeps its SSH connection. * fix(native-chat): one send per chat, with no line of sends behind it A chat with a send out took further messages into an in-memory line and sent them one by one. A message typed behind one in doubt then hit its own 30 s deadline and came back as not sent without ever going out. Now, as the common pattern does, a chat takes one send at a time: while it is out, Send is disabled and Enter leaves the text in the box. The sender refuses a second send instead of lining it up, so the 30 s deadline always runs from the send itself. A Stop or a tab close stops the one send: settled from its answer if its request is out, handed back as unconfirmed between attempts, or silently if it never went out. Notes sent from outside the chat while its send is out stay with their sender (not ready); a launch prompt that meets the person's own first message waits in the composer instead of being lost. * fix(native-chat): give a Stop-withdrawn note back to its chat once its notes were cleared Notes sent from outside a chat clear once their message is recorded. A message the host recorded as pending and a Stop then withdrew came back only to its sender, which had already let go of it, so the text was lost. The chat's draft now takes it, as an earlier build's outbox did. * test(native-chat): type the composer-actions probe without a cast * chore(native-chat): drop outbox wording left in comments and an empty locale group * fix(native-chat): pace failed checks, keep the host's reason, and hold the chat for its launch prompt - A send whose checks failed before its request went out (an unreachable or incompatible host, an unreadable history) was tried again at once, about a thousand times a second for 30 s. Attempts are now paced by the attempts made, whether or not their request went out. - A host that refuses every resend by throwing (native chat turned off, a journal that won't open) gave back only "couldn't confirm". The host's reason now comes first, still without claiming not sent. - A launch's prompt now holds the chat's one send from the click, drawn as sending, so a message typed while the chat starts can't overtake it; it goes out once the chat exists, and a cancelled launch frees it. - Notes whose send nobody could confirm say so instead of "did not accept", and notes launched into a new chat let go of their text once that chat's composer holds it. - An open chat keeps drawing a recorded send until its row arrives, so a reply that beats the history no longer makes the message flicker out. * fix(native-chat): hold a chat's sends until it has started, and send a failed chat's message with its restart A message sent to a chat still starting went out at once to a host that had no record of the chat yet, read for a fence it could not get, and came back as not sent. A message sent to a chat whose start failed restarted it but no longer went with the restart. While a chat starts, Send stays off and Enter leaves the text in the box, as with a send already out. A text message sent to a chat whose start failed restarts it and goes as the restart's first message, through the same staged-prompt path a launch prompt takes: it holds the chat's one send slot, drawn as sending, until the chat exists, and comes back to the composer if the restart fails again. Notes wait while a chat starts and ride a failed chat's restart, keeping their text until it is sent. * chore(native-chat): test the queue request, rejection words and card hand-offs; drop an unused clear - Pin which sends ask the host to queue, the moved rejection wording, and that a queue send whose card was handed off and then refused or withdrawn, or whose replay names a withdrawn card, is never handed back. - The send-at-most-once gate now says what the desktop promises after a reload: it never sends the id again, and hands the text back. - The legacy read names when it goes, and drops the notice clear nothing called. * fix(native-chat): keep a sent launch prompt's entry, and give back at once what can't go out - Cleaning up a launch prompt released its send slot even after the prompt had gone out, which deleted the entry the sender keeps once the host records it. A launch prompt recorded as pending and then withdrawn by a Stop was lost from both the chat and the box, and an open chat dropped the new chat's first message before its row arrived. The slot now gives back only a reservation that was never sent. - A send stopped before its request went out by something trying again won't clear (this client and the server can't talk, or the host refused the history read) kept Send off for 30 s and then said Orca couldn't reach the agent. It now comes back at once with its own cause: not sent, since nothing went out, or unconfirmed if an earlier attempt did. Transport errors keep the paced retries. * fix(native-chat): notes keep their own text through a new agent's launch Notes sent to a New agent whose start failed came back to the notes and also sat in the new chat's composer, so sending both delivered the text twice. The notes keep their text until it goes out (they hold it from the click), so the launch now says so and no composer gets a copy, on a failed start or a refused prompt alike. The tests that asserted the copy in the composer pinned the old double ownership and now assert the notes are its only owner. Notes that rode a failed chat's restart and ended unconfirmed now say Orca couldn't confirm them, as notes sent directly do, instead of that the agent did not accept them. * chore(native-chat): the send-once gate says what the desktop keeps across a reload The desktop keeps a send's id in memory only: a send still unsettled at a reload or crash is not resent and not handed back. The coverage notes no longer credit the renderer tests with durable identity across a remount. * fix(native-chat): say once why notes sent to a new agent did not go Since notes keep their own text through a new agent's launch, a prompt the host refused, nobody could confirm, or that found the chat's send taken came back to the notes with nothing said anywhere: the chat shows no notice for text its caller keeps. The notes menu now reports it once, with the toast a send to an existing chat already uses: not accepted, couldn't confirm, or not ready. A start that failed still says so in the new chat instead. * fix(native-chat): retry a history refusal the host says clears, and name it at the deadline A send stopped before its request went out came back at once for any refusal of the history read, including ones the host names as clearing (its journal briefly unavailable, a chat detached while the host quits), so the automatic retry was lost. Only a version mismatch and refusals that won't clear come back at once now; the rest go again on the paced schedule, and if the deadline still finds nothing sent, the words are the host's refusal rather than Orca couldn't reach the agent. * feat(native-chat): a send makes one request, and nothing ever sends it again The desktop resent a message under its own id for up to 30 s when the answer was lost. A send now makes exactly one request, as the common pattern's clients do: - An answer that is lost, dropped or proves nothing hands the text back at once with "couldn't confirm… check the chat". - A failure before the request goes out (the environment check, the history read for the fence, a refused connection) means nothing went out: the text comes back at once with its own reason, or as not sent. - The 30 s cap stays on the one request, so a host that never answers can't hold Send. Nothing ever resends, so nothing can go out after a Stop: a Stop only takes back a send still in its pre-send checks. The resend state goes with it (tries, generation, awaiting, stopped, the resend timer, the send-answers-proof probe, and the first-attempt/resend split in the evidence), and the send-once gate says the desktop never resends. * fix(native-chat): notes keep the chat's line, and a send held behind a rewind says it was not sent - Only a send whose text belongs to the chat's composer clears the chat's line. Notes sent from outside the chat no longer wipe a "couldn't confirm... Check the chat" line that explains text already back in the box. - A send the host turns away behind a rewind it could not confirm went back as "not sent" but said "couldn't confirm what happened. Check the chat". It now gives the rewind's reason and says the message was not sent. - Reliability gate names the single-request test and records a fresh evidence run; the sender test drops its leftover resend mocks and a duplicate Stop test. * fix(native-chat): a send ends even when putting its text back fails If writing the returned text into the chat's draft threw, the send never settled: it stayed "sending", the chat refused every later send until a reload. The send now always ends after a hand-back, the failure is logged, and the chat's line adds "Couldn't save your message." * fix(native-chat): a message typed during /clear stays in the box and follows the chat A /clear moves the chat to a new conversation, and its host refuses any send while it runs. A message sent then went out, came back refused into the old conversation's draft with its line, and vanished when the chat moved on: the box and the line now belong to the new conversation. - A /clear holds the chat's one send slot while it runs: Enter does nothing, Send shows busy and the text stays in the box, as for a send that is out. Notes sent from outside get "busy". - Once it moves the chat, the old conversation keeps taking no send until the view leaves it, and what is left of its draft moves into the new conversation's draft, after anything there: when the composer's /clear settles, and again when the view moves. * fix(native-chat): no "Send message?" or queue clear while the chat's send is out While a send is out (or a /clear runs) the chat takes no message, yet Enter over a held queue still opened "Send message?", and Clear queue deleted every card before its message was refused. Enter now does nothing there and the text stays; Clear queue re-checks and deletes nothing if a send went out after the question opened. * refactor(native-chat): take the /clear draft carry out of this change Moving the old conversation's draft into the one a /clear replaces it with fixes a bug main has too (text left in the box during a /clear stays with the old conversation), so it goes in its own change. Kept here: a /clear holds the chat's send slot while it runs, and the conversation it moved away from takes no send until the view leaves it. * fix(native-chat): a /clear's hold on sends always ends - A view that unmounted while a /clear that moves the chat was out left the old conversation holding its sends until a reload: the late reply kept the hold for a view that was gone. The reply now releases it. - The local call for a conversation command has no deadline of its own, so a /clear that never answers kept Send off for good. The hold now also ends at the command's deadline (195 s, the one the remote call already uses), whichever comes first. * test(native-chat): opening a chat an older build left stuck; hand back its copy in send order Pins, through the real chat hook, sends, legacy recovery and draft store, what opening such a chat does: the queued messages behind a message the host recorded in doubt come back to the composer once, in order, with the "not sent" or "couldn't confirm" wording; the chat is free to send under its read's fence; nothing happens while the chat has no fence here. The legacy reader now hands entries back in the order they were sent (queuedAt), as the older build's reader did, instead of array order. * fix(native-chat): a send this window turned away for a re-paired server comes back as not sent A managed server's update rotates its pairing, and this window's main process then answers the next call itself, before forwarding anything, with runtime_environment_changed. The send read that thrown answer as proving nothing, so the person was told Orca couldn't confirm a message that never left. It is now handed back as not sent, with the reason. Every other thrown answer still reads as unconfirmed once the request may have gone out. * Save a new chat's first message during create * Align opening-message types with the user-message wire * Keep chat publication and cleanup outside store initialization * Complete the create refusal fixture ledger * test(native-chat): prove a created first message waits for the agent's start and is handed over once * refactor(native-chat): keep the host and session hook under the line limit after the main sync * test(native-chat): a chat created with its first message and no pick still teaches its agent's default After the main sync, a created chat's agent starts from the delivery loop rather than the create call. Grok (through the real ACP adapter and model catalog) and Claude (through the real adapter) each still report the model their own config resolved to once the chat reads its options. |
||
|
Brennan Benson
|
a747c1c013 |
Allow another phone send after delivery is unconfirmed (#26392)
* Allow another phone send after delivery is unconfirmed * Update phone image recordings after journal removal |
||
 Neil
|
f4c7121545 |
Use browser crypto to speed up mobile image fingerprints (#26170)
* perf(mobile): use browser crypto for image fingerprints * Document web image fingerprint platform override * fix(mobile): share the image fingerprint domain |
||
|
Brennan Benson
|
2f49377425 |
feat(native-chat): Grok as a structured chat over the Agent Client Protocol (#25225)
* Leave a stopped turn's running tools to the agent's own end
When another writer settles the open turn (a person's Stop), the assembler
now only stops that turn's text and cancels its pending prompts. Running tool
calls stay the agent's: a progress update or completion it reports after the
Stop lands as reported, and whatever is still running settles at the agent's
turn end for that turn, the next turn's open, or the session's end.
An agent's end for an earlier turn while a newer one is open no longer clears
the open turn's activity line or ends its anonymous reply. An unnamed end right
after a Stop ends the stopped turn instead of being dropped. The test rig's
restart no longer writes the dead assembler's window text, matching dispose.
* Pin that a stopped turn's running tools hold budget until the agent's end
* Type the stopped turn's tool progress update as a tool body
* List every event the assembler hands to the decision step
The type-aware lint requires an exhaustive switch with no default case.
Also retitle a Stop test to say what it asserts.
* End a running call as its turn's journal row ends
A call still running when its turn ends takes the state of that turn's
row: a row another writer settled first (a person's Stop) stands, so its
calls read interrupted whatever the provider's later end reports. The
no-ending path that settled calls from the Stop row is gone, since a Stop
now leaves running calls to the provider. Adds the two Spanish strings.
* Say why a Grok turn failed, and keep task rows in Grok's own words
A failed Grok turn ended with no reason on screen: the translator dropped
every copy of Grok's message. The failed turn now gets one status row in
Orca's existing "provider did not accept this message" words with Grok's
reason, read from whichever copy arrives first (the given-up retry, the
turn's end, the prompt's completion notice, or the prompt's error answer);
later copies only fill a reason the row still lacks.
A running background command no longer reads "Background task <id>
started": a task's summary is mapped only once it has settled. A monitor
stays a monitor when the agent reads its output: a frame that names no
kind keeps the known one, and a "[monitor" command is a monitor.
A prompt's turn is marked started, so a late frame for an ended prompt
neither reopens it nor becomes the active turn. A tool's turn is held in
one place at a time.
* Read a monitor from Grok's exact output prefix
* Word a failed Grok turn in Grok's own text, not as a refused message
A turn that started and then failed was told "The provider did not accept
this message", Orca's sentence for a message refused before its turn. The
row now reads as a Codex turn-ending error does: an error status row with the
provider's own words. With no words, the dialect names the failure ("Grok
ended this turn with an error." / "Grok usage limit reached."), else the
agent's display name does.
* Settle a stopped turn's running call as its turn row ended after a restart too
The restart sweep ended every running call by the death evidence alone, so after
a person's Stop with no proof the child died the call read failed under a turn
that read interrupted. The sweep and the live dead-generation settlement now ask
the same rule the assembler does: a call in a turn already settled ends as that
row ended; only a turn still running leaves its calls to the evidence.
* Keep the dead-generation settlement under the line cap
* Register the ACP schema verify step in the PR preflight phase test
* refactor(agent-session): one required agent registry; declarations admit what they claim
A StructuredAgentRegistry, built once from the {definition, adapter}
registrations, is now a required host dependency and the router routes with
it. The adapter interface loses its optional router-only capabilities?() and
definition?(); every reader (options at rest, thread goal, rewind, the
/compact handover) asks the registry. A live session still narrows rewind
through the adapter, and the host combines declared and narrowed in one
helper. The registry refuses a registration that declares compact, a thread
goal or rewind without the adapter method behind it.
/compact is admitted by the declared capability, so an agent that declares
compact:false gets the commandRefused fact instead of a thrown error.
Also: the cut-turn notice names the agent from the catalog, create-support
builds the account home through agentSessionAccountHome, and the turn
status text older clients read names the session's agent instead of
defaulting to Codex (Claude/Codex text unchanged).
* Read ACP permissions, session events and prompt errors through the protocol client's own types
The translator now reads a permission request with the client's lenient reader, a session update
with its session-event reader, and takes only the agent's own error answer as a failed prompt's
reason, so an Orca-side error never reads as the provider's words. Tests cover protocol values
newer than this build.
* refactor(agent-session): the router applies the declared rewind itself
The router holds the registry, so its rewindSupport answers the owner's
declared rewind narrowed by the adapter, in one place no reader can bypass.
The host-side combining helper is gone; readers ask the adapter they hold.
* test(agent-session): register the agents the merged-in tests now need
The adoption replay test builds its host over this build's registered agents
instead of an adapter definition method, and the stored-form test passes the
stored agents the record guard now requires.
* chore(agent-session): the registry is the one lookup; drop the test-only empty capability record
* fix(agent-session): keep "dismiss all" restart offers dismissed for the desktop
The desktop renderer calls the restart methods as a paired runtime client
without the registered-agents capability, so it was handed a Claude/Codex
audience and its "dismiss all" took the scoped path: no dismissal fence and
no clearing of unwritten teardown witnesses, so a late teardown write could
bring a dismissed offer back.
The audience is now derived against this host's registry: a client that can
show every agent the host registered gets none, and dismisses exactly as
before (one fence for every offer). A client that truly cannot show some
agent still gets a final dismissal for what it sees: each cleared session is
fenced on its own (bounded, superseded by a dismiss-all fence) and only the
witnesses of agents it sees are dropped.
One predicate now answers which agents a client renders, for both tab
projection and restart offers; the Claude capability rule is folded in.
* fix(agent-session): a changed agent definition never hides that agent's chats
A record was readable only if every handle used the transport its agent's
definition declares today, and its account variable was declared by some
registered agent. A later build that drives the same agent over another
protocol, or under a renamed variable, would have set every existing chat of
that agent aside: no tab, no history, no error.
Readability now asks only that the record agree with itself: its agent is
registered, every handle shares one namespace owned by that agent, and the
account variable is a well-formed name. Whether this build can drive it (the
chain's transport is the one the agent speaks, and the account variable is
that agent's own, since it becomes the child's environment) is decided in
performAttach, the one admission every start of every agent passes through,
and refused there as hostUnsupported. A record pinning another agent's
variable is refused the same way rather than passing because some agent
declares it.
* refactor(agent-session): each agent's registration says where it runs and which account it pins
createSupport, create and the model catalog's account read still chose a
location rule and an account resolver by name (Claude, Codex, else none),
so a new agent would have needed a third branch beside the registration
list that already decides storage, routing and publication.
Each runtime registration now carries supportsLocation(location) and
resolveAccountHomePath({ launchEnv, location, purpose, workspacePath }),
and Claude's and Codex's rules move into their entries unchanged: a read
still syncs no home and starts no bridge, a Codex launch still trusts the
workspace first, and WSL locations resolve as before. The list exists
before the host is built, so none of these reads installs the host, and an
agent the list does not hold is answered no without opening the journal.
* test(agent-session): drop a duplicate registry key; keep fences out of the capsule state
A spread input already carries the registry, so the explicit key was
overwritten (TS2783). The session-fence helper now returns only the fence,
not the capsule state it was given.
* fix(agent-session): a scoped dismiss-all persists no per-session fence
The per-session dismissal fence added a forever-persisted capsule field
with an arbitrary cap that served no caller: only the local desktop calls
the restart methods, and on a host whose agents it all shows it gets no
audience and takes the unscoped, fenced dismissal. A scoped dismiss-all
(a caller that cannot show some registered agent) keeps that agent's
offers, clears only its audience's unwritten witnesses, and writes no
fence; a late write from this process is serialized behind it. The field
never left this unreleased branch.
* fix(agent-session): refuse an attach whose agent is not the session's own
The start check judges the record's agent while the router starts the adapter of params.agent, and the attach wire let the two differ. Admission now refuses agent !== provider as requestMalformed, so the agent checked is the agent started. Every host-built attach already sets them equal.
* fix(agent-session): offer to start a chat only when the start would accept it
The restart offer, a failure's retryable flag and the pre-send check asked only whether the adapter runs the chat's location, while the start also refuses a record this build cannot drive; such a chat was offered Resume and its Retry failed forever. One host predicate, hostCanStartRecord (location support and agentDrivesSession), now answers all four; adapterSupportsRecord stays the reader gate, and an undrivable chat's offer is kept, not retired.
* fix(agent-session): the model-catalog probe uses a record's account only if this build drives it
A record-scoped catalog read started the agent's lister under the record's account path whatever variable the record pinned. It now uses that path only when agentDrivesSession holds, and otherwise resolves the account as for a read with no record.
* refactor(agent-session): the record store admits agent ids; comments say where transport is checked
The store only ever asked whether an agent is registered, so its admission list is now the registered ids; the definition is the one home of an agent's transport and account variable. Comments that said the record store checks transport, and one that cited a nonexistent function, now point at agentDrivesSession and the attach admission. The isPersistedAgentSessionRecord(value, agents) call shape and the test fixture the cross-version probe imports are kept.
* docs(agent-session): the record store admits the registered agents' ids
* refactor(native-chat): Grok's registration declares where it runs; ACP no longer borrows Codex's location rule
The rule a self-supervised agent child runs under (this machine, no WSL, Windows only with process
start-time proof) is its own module that Codex and the ACP adapter both use. Grok's registration
takes the full account-home resolver signature, and D3's tests build hosts with the agent registry.
* fix(native-chat): Grok follows the ACP runtime's request contract and the managed process's close
A request the agent or a Stop cancels is answered with the agent's own cancelled reply by the code that
owns it (the runtime no longer answers a silent handler), so a Stop needs no separate decline pass. A
permission answer still being saved when the agent stopped waiting is reported unconfirmed, since the
protocol already answered it cancelled. Cancelling the agent's own turn is the plain cancel. Request
rows are matched under their generation-scoped ids. A refusal's reason comes from the dialect's wording
path. The child drops its own stderr tail and close policy for the managed process's, and a close
whose process tree was not proven gone is reported as the adapter contract asks.
* fix(native-chat): a Grok chat Orca already holds resumes without writing what Grok replays
A chat with a saved Grok session reattaches with session/resume where the agent offers it, else
session/load. Either way the call runs inside the translator's load window, so what Grok sends while
it reattaches (its saved exchange, a task the dead process left running, ended by the restart) opens
no turn and writes no row; only context usage reads on. A reply an Orca or Grok crash cut short is no
longer completed from Grok's saved history: it reads like a Claude or Codex chat's, with the existing
notice. The attach window also closes after a failed attach, and a created session that session/resume
reports missing is replaced like one session/load reports missing.
The replay reconciliation is removed: the lane no longer reads the journal, and D3's replayed-input
grammar test and completed-turn check in the assembler go with it.
* refactor(native-chat): a failed Grok reattach needs no window close of its own; its lane is replaced
* test(native-chat): D3's merged tests use the shipped declarations and the launch options main requires
* fix(native-chat): typecheck fallout of the base merges; any agent's empty chat is reusable
Main's idle-empty-chat lookup and launch join now take any registered agent, as the rest of the
launch path does. The refusal check moved into the prompt turns and the prompt-block conversion beside
the turns that send it, keeping both files in their line limit.
* fix(native-chat): a Grok Stop ends the process once Grok settles its turn; the next send resumes
Grok's session/cancel ends only the running turn: work it already moved to the background keeps
running and can begin a turn of its own after the person pressed Stop. Stop is now a session
boundary, as it is for Claude: the cancel answers open requests and lets Grok end the turn, the host
waits a bounded grace for that, then ends the process; the next send relaunches and resumes.
The adapter's own bounded close of a turn Grok began is gone. Its named-turn check stays: the host
ends the session unless the provider declines a Stop naming a turn that has since ended.
* test(native-chat): a Grok Stop ends the process only after Grok answered the cancel
* fix(native-chat): Steer on a Grok card cancels the running prompt, then sends it
A send that reached Grok while a prompt ran was held in the adapter until that turn ended: Steer
on a queued card took the card out of the host's editable queue and meant 'send after this turn'.
It now cancels the running prompt (session/cancel; the session stays) and sends as the next prompt
once Grok answers the cancel, as the common pattern does; a steer behind another cancels it in
turn, so the last one runs. The adapter holds a send only while that cancel lands, so its general
held-send queue and its holdsDispatch report are gone (every send it holds has its turn open in
the journal). An older client's mid-turn send takes the same path. capabilities.steering is
unchanged and still unread.
* refactor(native-chat): a close or Stop cancels a start through the acquire's own abort signal
The host owns the acquire it runs, so it now owns its cancellation: each attach's acquire gets an
AbortSignal, aborted from outside the session's queue by a close and by a Stop admitted now (the
same admission rule as before). The optional abandonStart adapter hook, the router's fan-out to
every adapter and the ACP adapter's session-keyed start map are gone; the ACP adapter keeps an
unkeyed set of starts only so quit can prove their children gone, and keeps a failed start's
unproven child until its exit is proven.
The hook also let a later close ask that child again. The host now does that from state it holds:
a close of a chat with no live child whose record still names an owner process with no death
evidence asks the adapter to release it. The answer is not recorded as proof (the lease probe
does that), so an owner pid an earlier Orca left is never killed or marked gone. Claude and Codex
ignore the signal and hold no such child; their release is a no-op (tested).
* fix(native-chat): a Grok crash that closes stdout before its exit still ends with Grok's last words
On macOS and Linux the agent's stdout ends before its exit is observed, with or without the
supervisor's EOF forwarding, so the connection's loss closed the journal first and its error text
became the session's ended reason, dropping Grok's stderr. The reason is now read at the proven
exit: the agent's last words when it left any, else why the connection closed. The failure already
carried them. Comments that assumed the exit comes first, that early frames past the cap refuse the
start, and that dispatch re-checks image support are corrected.
* fix(native-chat): nothing Grok sends while a held chat reattaches is written, marked as replay or not
The reattach window relied on the dialect's replay verdict, and Grok's frames read as live unless
they carry isReplay, so an unmarked chat frame during session/resume opened a turn that never
ended. D3 now marks every frame inside the window as replay before the translator reads it, so the
translator keeps only context usage whatever the agent marked; options and commands are still
adopted. The translator's load semantics are unchanged.
* test(native-chat): a Stop after a resume finds no turn an unmarked old reply opened
* test(native-chat): a resumed Grok chat keeps its last context reading; the resume refreshes only the window
* test(native-chat): a Grok background task a Stop ended reads as stopped reporting
* refactor(native-chat): quit's stop of each start answers through one promise kind
* fix(native-chat): quit aborts every start the host has in flight before draining attaches
A Grok that never answered its handshake held quit until the start's own 60 s bound, past the
20 s quit deadline. The host's teardown now aborts each in-flight acquire (and any the drain
still begins), so the adapter's own quit controller and its map of starts are gone: a start
has one canceller, the host's signal.
* fix(native-chat): a Grok start's abort stops reaching its child once the start has returned
The listener stayed on the host's signal until the attach finished committing, so a Close in that
window killed the now-live child behind the host's back and it read as Grok crashing. The start
now detaches it when it ends; a later Close goes through the session's own stop.
* fix(native-chat): a close or Stop during any attach phase stops the start before it launches
The attach began its abort controller only after reconciling leases, resolving recovery and
probing the previous owner, so a close or admitted Stop in those phases reached nothing and Grok
launched anyway. The controller now begins first, and the acquisition checks it before asking the
adapter to start.
* test(native-chat): a close during the attach's owner probe asks no adapter to start
Also renames the close test after the hook it no longer exercises.
* test(native-chat): a close's re-ask closes a Claude or Codex child a failed cleanup left
The re-ask is not a no-op for them: when the adapter still holds the child its cleanup could not
prove gone, the close stops it again as a requested close, and Claude persists the handle of the
conversation it ran so the next send resumes it. Corrects the tests' and comment's wording; the
close awaits the re-ask, bounded by each adapter's kill ladder.
* fix(native-chat): Steer during a turn Grok began itself cancels it and sends once it ends
A send while Grok ran a turn of its own (a background task waking it) went straight to Grok, which
queued it behind that turn where Orca could no longer withdraw it, while Stop treated the same turn
as the running reply. The send now waits as a steer, the turn is cancelled once, and the message
goes when the turn ends; a Stop withdraws it and an exit rejects it as never sent.
* test(native-chat): a steer whose cancel Grok never answers ends Grok and is rejected as never sent
Pins the bounded steer cancel kept from the runtime: past the bound the connection closes, the
running reply reads unverifiable, Grok's end reads as its exit, and the waiting steer is rejected
as never sent.
* fix(native-chat): a Grok crash stays a crash when a stop lands before its exit is proven
After the connection broke and the close could not prove Grok's exit, any later stop Orca asked
for (the next start, a Stop, a Close) marked the child as closed by Orca, so the crash read as a
requested close and Grok's last words were dropped; a send meanwhile was recorded unconfirmed.
The connection loss now decides the cause, and a send on that session is rejected as never sent.
* test(native-chat): fixtures this PR's registered Grok and desktop capability made stale
CI's unit shards failed on tests outside the PR's own lists. Each encodes something this PR changes
on purpose: Grok is now a registered agent (the seam test's unregistered agent is now Cursor); the
desktop now advertises registered agents (the restart-offer tests' older client drops that
capability explicitly); the attach context carries the start's abort controllers (the forget-status
double gains them); and the ACP real-host test rig sends to the host directly (listed beside the
other real-host rig in the send ratchet).
* fix(native-chat): a start quit stops is not the queued message's start failure
With quit now aborting a start it would have waited for, the delivery step recorded the aborted
start as the message's failure ("couldn't restart"). After quit has stopped delivery, the step
leaves the message to quit, which settles it as a close does ("The chat closed before this message
was sent."). The test that pinned quit waiting for that start and stopping its child now pins that
nothing is launched behind quit.
* fix(native-chat): a message sent after a Stop or close aborted a start gets its own start
A start the host aborts (an admitted Stop, a close, or quit) returned its refusal to the delivery
loop, which then rejected whatever was queued at that moment with "couldn't restart", including a
message the user sent after the Stop. The attach now reports that the host aborted it, and the loop
re-derives from the journal instead: what the Stop or close withdrew is already settled, a message
accepted since gets a start of its own, and quit's next step stops the loop. This replaces the
quit-only carve-out with the same rule for every abort and every agent.
* test(native-chat): the message sent after an aborted start is answered, so no settlement outlives the test
* fix(native-chat): a Grok model pick Grok never answers no longer holds Stop or Close
The pick runs on the session's queue. It now registers in the host's out-of-queue
abort registry beside a start, so a close, an admitted Stop or quit abandons it, and
the ACP adapter bounds it at 30 s like Claude and Codex. A late answer is still adopted.
* fix(agent-launch): a phone's launch opens a terminal for an agent whose chat it cannot show
agent.launch now reads the caller's capabilities by the rule tabs and restart offers
use (clientRendersStructuredAgent). A phone without registered-agents.v1 gets Grok as
a terminal again, as on main; the host's own callers and desktop clients are unchanged.
* fix(acp): strip every agent hook variable from the ACP child, from the shared list
ACP_CHILD_ENV_TO_DELETE was a second copy of the hook runtime keys that missed
ORCA_AGENT_HOOK_TRANSPORT; it now spreads AGENT_HOOK_RUNTIME_ENV_KEYS beside the pane
identity keys.
* refactor(native-chat): the mutation context carries the provider-wait registry itself
Keeps the host file within its line limit; one field instead of two closures over it.
* fix(agent-launch): agent.launch.v2 still vouches for Claude and Codex chats
The caller rule from the previous commit also turned Claude and Codex into terminals
for a client advertising only agent.launch.v2, whose contract says it opens a chat
(mobile retry-authority tests). Only an agent beyond those two now needs the client to
read it (clientRendersStructuredAgent); the test fixtures go back to what they were.
* refactor(native-chat): drop saved-history adoption from the timeline assembler
The common pattern discards the history a provider replays while loading a
session, so the assembler has no use for an input.history event.
* refactor(acp): drop session/load history adoption from the translator
The common pattern discards the history an agent replays during session/load,
keeping only what it says about the context window. Remove the adoption path
(acp-history-adoption.ts, the adopt option, and the historical background-task
liveness rewrite it fed) so load replay is always dropped except usage.
* refactor(native-chat): a pending input is only Orca's send now
Review follow-up to the adoption removal: drop the comment naming the
provider's saved message, and make requestedAt required since every pending
input comes from input.accepted.
* test(acp): keep the task-result status table on live frames
Review follow-up to the adoption removal: the result-status mapping was only
tested through adopted history, so run the same table on live frames, and
cover an unmarked task notice during a load being dropped.
* test(acp): a frame helper for a shell command Grok is running
* fix(acp): a Grok crash settles through the host's provider-exit batch, scoped to the turn it ended
A Grok crash ended the journal unverifiable before the adapter reported the exit, so the host's
provider-exit settlement found no running turn and wrote nothing: the adapter's failure (with
Grok's last words) never reached the journal, and a later stale-session pass wrote a bare,
thread-scoped cut-short row, so the partial reply was not folded as Claude's and Codex's are.
At a proven exit the ACP lane now ends its running turn interrupted at the exit instant, as the
host's exit contract expects of a child's own translator (Codex's does the same). When Grok's
stdout closed first (every POSIX crash), the turn is unverifiable only until the exit is proven:
the host's provider-exit settlement now takes the exit as proof naming the child's fence and
revises what that child left unverifiable in the same batch, with the turn-scoped row and the
adapter's failure. Claude and Codex write no unverifiable turn of a live child except a command
whose hand-off is in doubt; that turn is now revised at the exit instead of at the next open.
* test(acp): a crash seen first leaves the host no Grok turn to revise
* refactor(native-chat): what a gone generation left unfinished gets its own module
The settlement file passed 300 lines with the exit-proof revision. The unfinished-work reads
(capture, interrupted-by-the-exit, in-progress) are their own concept and move out unchanged,
apart from the exit proof they now take.
* refactor(native-chat): a watched exit revises what its child left unverifiable without reading Stop marks
An exit's own instant is the turn's end, so the revision needs only each row's fence: the
settlement's journal type gains itemFence alone, and the host test fakes say so.
* test(native-chat): drop the duplicate itemFence on the fake that already had one
* test(claude, codex): an exit whose stdout ended first still reports as it always did
The provider supervisor now ends Orca's stdout when the agent's ends, so on every crash EOF
arrives before the exit is seen. Claude's and Codex's connections report nothing at EOF and
report the exit, with its usual reason, once it is seen.
* fix(acp): reopen a chat with session/load, as the common pattern does
An agent that offers both now reloads its session instead of resuming it; the
reattach window still discards what it replays except context usage.
* fix(acp): drop the 60 s handshake bound; an abort fails the start's waits at once
Neither common design bounds an ACP handshake: Close, Stop and quit end a start
that never answers. The abort now also closes the connection, as a kill there
does, so the start settles even before the child's exit is proven. The
host-stopped start refusal only this bound produced goes with it; the idle
sweep keeps its words.
* fix(acp): a Stop naming an ended turn follows Claude's rule
It still stops nothing while another turn is live, but in the gap before a
follow-up's turn opens, which no client can name, it now stops what is in
flight and the session ends, as a Claude Stop does.
* fix(native-chat): a close no longer re-asks a failed start's unproven child
Neither common design retries that stop at Close, and Orca's Claude contract
re-asks only at the next start and at quit. The ACP adapter keeps the child
until its exit is proven and asks it again there, as Claude does.
* fix(acp): a message sent during a turn the agent began itself goes at once
Both common designs send it straight to the agent with no cancel; only Orca's
own running prompt is steered (cancelled, then re-prompted).
* test(native-chat): dismiss-all through a remote client's audience keeps a newer Orca's offer
Uses an audience production sends (one that cannot show every agent), per review.
* fix(acp): launch Grok as `grok agent stdio`, without the update and leader flags
The common pattern passes neither --no-auto-update, --no-leader nor
GROK_DISABLE_AUTOUPDATER; full access still adds --always-approve.
* fix(acp): an agent that ends its stdout, or answers unreadably, is not a lost connection
As in the common pattern, only a broken stdin (or Orca's own close) ends the
agent; one that closed its output but can still be written to stays until a
Stop, a close or its exit. The provider supervisor goes back to its base
content, so Claude and Codex no longer get the forwarded stdout end either.
* fix(native-chat): a person's close joining a failed one still binds the turn its child end cuts
On main every close of the chat writes its own Stop and settle. Here a later close joins the
first and writes no row, and the first's settle closed when its kill failed, so a turn that opened
in between and was cut by the next close read as failed. A person's close joining a person's close
whose Stop opened a settle now reopens that settle until its attempt is done.
Tests: a close whose kill failed still closes its settle; a turn opened between a failed close and
the next reads as the person's cancellation (each fails without its half of the fix).
* test(native-chat): Grok opens as a chat only behind the structured-chat setting
agent.launch and orchestration worker-start read the same setting as the
renderer route; pin both states for Grok on each. The setting's description no
longer names only Codex and Claude, in every catalog.
* docs(acp): generic ACP comments say what holds for every agent, not Grok
Stop ends the session for every ACP agent, as in the common pattern; the
adoption hook comment goes (adoption is not planned); a failed start's child is
retried at the next start or quit.
* test(claude, codex): type the EOF-before-exit test's streams; the supervisor no longer forwards EOF
The Claude test wrote to the child's stdout and stderr through their Readable
type, which the node typecheck rejects; it now holds its own PassThrough
streams. The comments no longer credit the reverted supervisor change.
* feat(acp): a steer's cancel asks once and never ends the agent
The runtime had one cancel: send session/cancel, wait at most 10 s for Orca's prompt to settle,
then close the connection, which ends the agent. A steer used it too, so a slow agent lost its
process just because the person added a message. requestSteerCancel() now sends session/cancel
once per prompt, cancels the agent's open requests and answers later permissions cancelled, and
never bounds or closes: the prompt's own reply ends it and the steer's prompt follows. cancel()
stays the Stop: bounded, then close. A Stop after a steer still bounds and closes. Both cancel
paths move into acp-prompt-cancel.ts over one cancel channel.
* chore(native-chat): keep the record store and recovery capsule under max-lines after the main merge
* fix(acp): a repeated steer shares the cancel in flight; say what the caller owns
Per review: a second steer before the first write lands returns that write instead of resolving
early. The steer's JSDoc says the wait for the prompt's reply is unbounded and that a prompt that
fails instead must not take the steer until the caller rebuilds the session; the Stop's says a
prompt that settles in time leaves the agent for the Stop's owner to end. The steer test now gives
the runtime a handler that would allow: the open permission's signal aborts and the late one never
reaches it.
* fix(native-chat): drop the stopDelivery the A3 merge doubled
* fix(acp): a steer's cancel asks Grok once and never ends it
A steer now uses D1's notify-only cancel. Two messages sent during a reply Grok began itself
cut that reply, as the common pattern does, and then both run; before, the queued first
message could not answer the bounded cancel and Orca ended Grok although Grok answered.
A Stop keeps the bounded cancel and its 4 s grace.
* fix(acp): a permission Grok asks with no prompt of Orca's running is declined
During a turn Grok began itself nobody asked it to act, so the request is answered
cancelled at once instead of opening a card that waits, as the common pattern does.
* fix(acp): a Grok that dies while starting is reported with its own last words
A dying process's stdout ends before its exit is seen, so the start failed as a closed
connection and Grok's stderr was lost. A start whose connection closed now waits, bounded by
the Stop grace (or a Close/Stop), for the exit before it is told.
* test: a Stop after a steer sends its own cancel; drop the import the A3 merge doubled
* test(native-chat): main's Stop-note test builds its turn context with the agent registry
* test(claude): say why the close test's fake child cast is safe
* test(native-chat): build the Stop-opened-turn test's identity and turn context the current way
The test (#25056) landed before the opaque provider handle (#24991), so main still built
the old {kind, threadId} handle; the turn context also needs this branch's agent registry.
* test(native-chat): build the Stop-opened-turn test's identity with the opaque handle
The test (#25056) landed before the opaque provider handle (#24991), so main still built
the old {kind, threadId} handle.
* test(ratchet): require src/main/provider-process now that it has landed
* test(native-chat): keep main's opaque-handle import in the Stop-opened-turn test
Main's #25706 and this branch both added the import at different lines; the merge kept both.
* test(native-chat): keep main's opaque-handle import in the Stop-opened-turn test
Main's #25706 made the same fix as this branch at a different line; the merge kept both imports.
* feat(native-chat): record a fresh provider conversation that replaced one the agent could not restore
A chat whose saved conversation the agent cannot reopen can now continue in a
fresh one: the handle chain records the new conversation as a creation that
replaces the lost one (which, why, and when), keeping every earlier link.
Rows keep a shape older builds read: the stored chain starts at the latest
replacement and carries the earlier links inside it.
* test(native-chat): build this stack's journal identities with main's opaque provider handle
Main's #24991 replaced the {kind, ...} handle with {transport, agent, nativeId}; three test
files from this stack still wrote the old shape. Same lines the downstream ACP branch uses.
* docs(acp): every reattach drops the agent's replay, not only for a chat the journal holds
* refactor(native-chat): store a replaced conversation flat; refuse it where older builds read the row
Older builds only read Claude and Codex records, so the nested stored form
protected rows no replacement can reach while adding a cap mismatch after a
downgrade. Store the chain as held, refuse a replacement in a Claude or Codex
chain until one has a stored shape older builds read, and refuse a supersession
key on a replacement that names no creation in the chain.
* refactor(native-chat): read hosts' structured agents from the app-shell services
Main grew the startup hydration hook to its line limit; the host agents sync is an app-lifetime subscription like the structured session tabs sync beside it, so it moves there.
* Use current provider handles in transition tests
* Use current provider handles in timeline fixtures
* test(native-chat): prove replacement rows survive downgrade and re-upgrade
* Require the ACP directory in the runtime import check
* test(ratchet): require src/main/acp now that this PR lands it
* feat(acp): a saved session the agent cannot reopen continues in a new one, with one warning row
When session/load (or session/resume) of a saved ACP session fails, the chat starts a new session and records it as a creation that replaces the lost one (#25747's 'replaces' link), and writes one warning row that the agent no longer remembers the earlier messages. A created session the agent reports missing is still superseded silently; a signed-out agent or a start that is over (Close, Stop, a lost agent) still fails the start.
* chore(acp): rewrap the acquire header comment
* test(acp): a start closed while the agent reopens fails without opening or announcing a new session
* Let ACP connections own their supervised agent process
* Preserve ACP cleanup evidence and isolate exit observers
* Expose ACP cleanup observations and type the permission fixture
* refactor(native-chat): the registered-agents capability lives in its own module
Main's growth put protocol-version.ts one counted line over its 300-line limit once the capability
was added; like main's other per-feature capabilities, it now has its own module, and importers read
it from there.
* refactor(acp): one connection owns the Grok process and its protocol
D3 now opens each ACP agent through createAcpAgentConnection (ACP-ALIGN #25810): one object spawns the
process on the execution host, owns its stdio and protocol, and reports its proven exit. It is built and
tracked before the handshake, so a start's abort (Close, Stop, quit) still reaches it, and a failed start
keeps that same connection for the next close to retry rather than spawning another process.
Deleted: the spawnAcpStructuredChild wrapper and its test, the raw-stream runtime assembly, the caller's
exit -> runtime.close wiring, the stdout-EOF heuristic (the connection no longer treats stdout EOF as
exit), and the 10 s steer/Stop cancel bound with requestSteerCancel. Reader control maps to
pauseReading/resumeReading; a close is connection.close after the host's existing 4 s Stop grace.
The adapter owns what the protocol no longer does: one session/cancel per running prompt however many
steers arrive (cleared with that send's settlement, retried after a failed write), and a Stop or steer
answers every open agent request the person has not already answered with the agent's own cancelled
reply. An answer already being saved when the Stop lands is sent.
Tests: blocked cancel write never holds Stop's grace, two quick steers send one cancel, a failed cancel
write is retried, a real process exiting while a child holds its stdout ends the session, and the
existing start-abort, retention, crash, connection-loss and reload-failure suites on the new rig.
* fix(acp): Grok signs in on its own machine with its API key or cached sign-in
When Grok reports that it needs authentication, Orca now names a sign-in method on the machine Grok runs
on, read from the same environment Grok was launched with: xai.api_key when XAI_API_KEY is set there and
Grok offers that method, else cached_token when Grok offers it, else none and the chat keeps the existing
not-signed-in refusal. The rule lives in Grok's launch spec; the adapter applies any agent's rule for new
and reopened sessions through the protocol client's caller-named method (authenticate, then retry once).
No new sign-in UI; interactive methods are never chosen.
* fix(acp): the adapter decides which of Grok's requests reach the person
The turn owner now admits every agent request, permission or question, from its own turn state: a
request reaches the person only while Orca's prompt runs and no steer or Stop is cutting it short (a
question may also come from a turn Grok began itself, until a Stop). Anything else gets the agent's
own cancelled reply and opens no card, so a question arriving after Stop or during a steer never
appears. A steer, like a Stop, withdraws the requests already open; an answer already being saved is
still sent. The protocol client's abort-on-cancel path is no longer used: after the connection
change its request signal aborts only when the connection closes.
* fix(acp): a plan Grok proposes shows as a plan, with no approval card
When Grok leaves plan mode it asks the client to approve its plan (x.ai/exit_plan_mode). Orca showed a
blocking 'Approve plan / Request changes' card for it; the common pattern has no such gate. Now the
plan goes into the chat's existing Plan row (the plan-document status row Codex and ACP plan updates
already use) and the request is answered at once with 'abandoned' plus feedback telling Grok to stop and
wait for the person's feedback or a request to implement it in a later turn, so nothing is approved on
the person's behalf. Dialects gain settleRequest for requests answered without asking anyone.
* fix(orchestration): worker-start opens a Grok worker in a terminal, as before
With the structured chat setting on, worker-start decided 'structured' for Grok and then the structured
worker factory (Claude and Codex only) refused it, so the start failed; main opened a terminal Grok
worker. Worker-start now decides with no registered agents beyond Claude and Codex, so Grok gets a
terminal worker as before. agent.launch and the app's own launches still open Grok as a structured
chat. Temporary until structured workers take registered agents.
* fix(acp): a prompt answer Orca can't read ends the turn instead of hanging it
A session/prompt rejection that was not the agent's own error answer (an answer that fails Orca's
schema, or one too large to read) left the turn running: the next message became a steer with nothing
to cancel and was never sent or settled, and Stop waited its full grace. As in the common pattern, any
prompt failure now ends the turn as failed (a failed-turn row without words, since none are the
agent's) and settles the send, so the next message goes. Only a closed connection keeps the send
running, for the connection-loss path to settle.
* fix(acp): send Grok's prompt-identity extension only to agents that echo it
session/prompt carried _meta {promptId, requestId} for every ACP agent, though only Grok's dialect
echoes it (injectedPromptIdentity). Now only an agent whose dialect declares it gets the extension;
other ACP agents get a plain prompt.
* refactor(native-chat): the registered-agents capability lives in protocol-version again, as on main
This reverts
|
||
|
Brennan Benson
|
3a03441580 |
refactor(native-chat): structured agents declare their capabilities instead of shared code naming Claude and Codex (#25076)
* refactor(native-chat): keep the provider resume handle opaque to shared code
Shared structured-chat code parsed each provider's resume handle: Claude's
session id and branch leaf, Codex's thread id, through a 'claude' | 'codex'
union every new agent had to widen. The in-memory handle is now
{ transport, agent, nativeId, providerData? }: shared readers use nativeId,
lease and handle-chain checks compare transport and agent, and only the
Claude adapter reads its leaf (providerData).
Stored and wire forms are unchanged for Claude and Codex. One encoding
module writes their typed shapes and decodes both those and the neutral
shape a new transport uses, which an older build refuses as unreadable
rather than reading as Codex. Key and root strings, which fork seeds,
superseded creations and resume offers persist, stay byte-identical.
The journal's own handle type becomes the journal-row and attach-wire
encoding of the same handle, and the journal identity carries the
neutral handle (null before the provider proves one).
No user-visible change.
* fix(native-chat): derive journal-row provider handles from the journal identity
The journal row converter now takes the identity every caller already holds,
so a row's handle has one obvious constructor. Tests that wrote the in-memory
handle straight into journal rows now build it through that converter, and the
processless Claude fixture names a not-yet-proved handle as null.
* fix(native-chat): refuse a stored provider handle written in both forms
A typed Claude or Codex handle that also carries the neutral form's
transport, agent, native id or provider data named two identities; it
was read as Claude or Codex and the next write dropped the other one.
Such a row now stays unreadable and is set aside untouched.
* refactor(native-chat): route structured agents through registered definitions
The structured-chat shared layer named Claude and Codex in a closed router
type, in option reads, in frame classification, and in renderer label and
validation branches, so adding an agent meant editing shared types. Each
agent now has a definition (capabilities, resting option rules) owned by its
own module; the router is a registry of definition + adapter pairs built at
runtime setup, and shared code reads the definition instead of the name.
No behavior change for Claude or Codex; no wire or stored shape change.
* refactor(native-chat): name the structured agent list once in host types
The host types, conversation commands, provider-session ownership and the
runtime's create path spelled the structured agent list inline; they now use
the one shared type the record and wire already validate against.
* fix(native-chat): narrow the record before reading its agent's option rules
* refactor(native-chat): make the router's registrations the only agent definition lookup
The at-rest option reads looked definitions up in a second, separately populated map, so a
registered definition could route and declare capabilities while the static map decided which
options a chat at rest accepted. The router now answers `definition(agent)` from its
registrations, the at-rest readers take it through the host's adapter, and the built-in
definitions are listed only where the runtime composes the router.
* test(native-chat): use opaque handle in queued rejection fixture
* test(native-chat): share one Codex journal identity in the integration suite
Main grew the suite to the 800-line limit; the opaque-handle import pushed it
over. The two tests built the same identity inline.
* refactor(agent-session): name the handle's adapter state resumeCursor
Rename the neutral provider handle's providerData to resumeCursor before any
row persists the neutral form: it is an adapter-owned resume position (Claude's
transcript leaf), never identity. Claude/Codex stored and wire bytes are
unchanged; their typed shapes never carried the field.
State the stored-form contract (a handle's field set is closed; later per-link
data goes on the chain link, which every build preserves) and pin it with a
record round-trip test. Document that transport records the id space the
native id was minted in, which can differ from the agent's current transport.
* refactor(agent-session): one required agent registry; declarations admit what they claim
A StructuredAgentRegistry, built once from the {definition, adapter}
registrations, is now a required host dependency and the router routes with
it. The adapter interface loses its optional router-only capabilities?() and
definition?(); every reader (options at rest, thread goal, rewind, the
/compact handover) asks the registry. A live session still narrows rewind
through the adapter, and the host combines declared and narrowed in one
helper. The registry refuses a registration that declares compact, a thread
goal or rewind without the adapter method behind it.
/compact is admitted by the declared capability, so an agent that declares
compact:false gets the commandRefused fact instead of a thrown error.
Also: the cut-turn notice names the agent from the catalog, create-support
builds the account home through agentSessionAccountHome, and the turn
status text older clients read names the session's agent instead of
defaulting to Codex (Claude/Codex text unchanged).
* refactor(agent-session): the router applies the declared rewind itself
The router holds the registry, so its rewindSupport answers the owner's
declared rewind narrowed by the adapter, in one place no reader can bypass.
The host-side combining helper is gone; readers ask the adapter they hold.
* chore(agent-session): the registry is the one lookup; drop the test-only empty capability record
* test(native-chat): build the Stop-opened-turn test's identity and turn context the current way
The test (#25056) landed before the opaque provider handle (#24991), so main still built
the old {kind, threadId} handle; the turn context also needs this branch's agent registry.
* test(ratchet): require src/main/provider-process now that it has landed
* test(native-chat): keep main's opaque-handle import in the Stop-opened-turn test
Main's #25706 and this branch both added the import at different lines; the merge kept both.
|
||
|
Brennan Benson
|
f0b3f44f10 |
feat(agent-session): let the host own a chat's tab id and let a create reserve it (#22616)
* feat(agent-session): let the host own a chat's tab id and let a create reserve it A structured chat's tab id was derived from its session id by every layer that needed one: the renderer, the host snapshot and the status address each built their own spelling. The join between a conversation and the tab that shows it must be a pointer the host owns, not a derivation each client repeats. The session record now carries surfaceTabId. A create pins it: the tab half of the pane agent.launch reserved, an optional tabId on agentSession.create, or a host-minted UUID. Records written before the field existed are backfilled at open with the string clients derived, in memory at once and on disk with the store's first transaction, so nothing keyed by it (read state, notification ids, worker rows) moves on upgrade. A second record under a held id is refused. Only the record and the two create wires change here. The snapshot still publishes agent-session:<sid> and the renderer still derives its local id; those move in the next two changes. agentSession.create is a strict object, so the field is advertised as a capability a client checks before sending it. * fix(agent-session): record the derived tab id for an unreserved create A create that reserved no tab minted a random UUID that no reader uses: the renderer, status address, worker rows and host-shared read state all still key by structured-agent-session-<sid>. Persisted, that id would move every chat created before readers switch to the recorded one, orphaning its read state and worker rows the way the backfill exists to prevent. An unreserved create now records the derived id, the same rule the backfill applies, so the record always matches the prefix every existing key uses; an opaque mint belongs with the change that moves the last reader. Also: - a chat tab id must be a host tab id on the record, the create wire and in admission, matching what agent.launch already requires of paneKey; a web-surface id would decode as another tab - the stored launch-result guard checks the structured outcome's tabId - comments no longer claim a retry naming another tab conflicts; replay keys on the attach fingerprint and answers with the recorded id (now pinned) - the wire refusal test used a non-hex digest, so the schema refused it for that reason; it now reaches the tab id rule - pin that the reload path refills the id without forcing a save * test(agent-session): correct the tab-id fingerprint comment to match replay |
||
 Brennan BensonandMerge Sim
|
8999a00281 |
refactor(native-chat): give each structured dispatch state exactly one meaning (#20133)
* refactor(native-chat): give each structured dispatch state exactly one meaning
`unknown` meant five different things. Only one of them was genuine
ambiguity.
A transport write that the provider's input pump never took is provably
undelivered -- which is what `rejected` already means. It was recorded as
`unknown` anyway, and a one-entry allowlist then existed solely to teach
Retry that this particular `unknown` was safe to re-deliver.
Collapsing that case into `rejected` deletes the allowlist and turns a
predicate into an invariant: Retry never re-delivers an `unknown`, with no
exception to reason about. The four states now each assert one thing --
`pending` written and awaiting, `accepted` the provider has it, `rejected`
provably did not happen, `unknown` genuinely cannot tell.
A fail-closed guard is the right default here because the asymmetry is
severe: refusing a legitimate retry costs the user a retype, while allowing
an illegitimate one sends the model a second copy of their message.
Also fixed, found while auditing every reader of `rejected`:
- The renderer printed `submission.reason` verbatim, so a broken pipe put
the internal token `provider_write_failed: broken pipe` on screen in
destructive red. The journal reason is unchanged -- it is the durable
evidence and the transport-versus-content discriminator -- but the screen
now gets copy that names the cause and says the message is safe to
resend. Content rejections still show the provider's own words.
- The fallback copy "Message was not accepted" read as a content refusal.
A null reason now yields "Message was not sent.", which asserts only what
every rejection shares.
- A refused worker-start preamble threw a plain Error out of the dispatch
path. It now throws `OrchestrationError('dispatch_preamble_undelivered')`
so a coordinator can tell "we could not send it" from "we sent it and
something else broke" without parsing prose. Retain/discard behaviour is
unchanged; only the verdict's legibility improves.
Two behaviours improve as a consequence rather than by design: a provably
undelivered message no longer blocks conversation commands, and no longer
leaves the session reading as "working" in chat and in every session list.
Not addressed here, and named rather than implied: a message left `unknown`
by a dead child or a host restart still has no recourse but retyping. The
restart reconciler that would decide those on evidence is written and has
never had a production caller. Parking the refused entry instead would
reintroduce the head-of-queue wedge removed in #19863, so it is not an
option.
Note for whoever edits `journal-reducer.ts` next: it sits at 297 of its 300
counted lines. The next statement added there needs a split, not a shave.
* fix(native-chat): close two gaps review found in the rejection taxonomy
Both are narrow and both were real.
A journal written before a refused write became `rejected` still holds that
submission as `unknown` with the transport marker. The predicate this change
replaced excluded exactly that shape from provider-echo matching; the
state-only check that replaced it does not, so on replay such a row could
claim the echo of a later, genuinely delivered send of the same text and
attach the delivery to the wrong message. Fail-closed still prevented any
re-delivery, so nothing duplicated — but the wrong submission was credited.
Replay now excludes the legacy shape too.
And the content-versus-transport split had a third case neither side covers:
a local capacity refusal is neither the provider explaining itself nor a
frame that failed to leave. It fell through to the verbatim branch, so
`claude structured dispatch queue is full` reached the screen — the same
class of leak this change set out to fix, one reason short of being caught.
Internal reasons now get copy; only a provider's own words are shown as
written.
Each is pinned by a test that fails with its guard reverted and passes with
it restored.
* fix(native-chat): preserve dispatch refusal across clients
* fix(native-chat): rotate immediately rejected retries
* docs(native-chat): correct rejection taxonomy reference
* docs(native-chat): align mobile retry comment
* docs(native-chat): clarify unknown replay semantics
* fix(native-chat): keep a mobile send's operation id when delivery is unknown
Mobile released the retained operation id whenever a send came back
`unknown`, so the user's next send of the same text went out under a fresh
id. A fresh id has no ledger row, so the host treats it as a first delivery
and dispatches it -- even though `unknown` is the one answer that says the
provider may already have the message. That is the duplicate this branch
exists to remove, reintroduced on the client that has no outbox.
Which case that was matters. Mobile only ever sees `unknown` from ack-loss
(`isRpcDeliveryUnknown`: "the host may have processed it and only the ack
was lost"), because the mapper reported every `ok` result as `accepted`
without reading `dispatchState`. So the rotation fired exclusively where
delivery was ambiguous and never where it was provably refused, which is
the inverse of the rule this branch establishes.
Retaining the id is what makes a retry safe, and it costs no liveness:
`performSend` answers a second request under a recorded id from the journal
and never puts it back on the wire, so a reused id delivers when nothing
landed and replays when something did. Rotating can only ever add a second
copy. The retention stays bounded by the host's admission window, which
`retainStructuredSessionOperationId` already enforces.
`retryUnknown` goes with it: the host ignores it for delivery, and all it
does is skip the cached answer to re-read the same row.
Keeping the id exposes what the rotation was hiding, so fix that too: a
replayed `unknown` comes back `ok`, and mobile called it `accepted` and
cleared the composer as if the message had landed. `dispatchState` now
decides, in one pure function:
accepted/pending sent, and the id is spent
rejected provably did not happen and terminal in the reducer, so
reusing the id could only replay that rejection: spent,
and the next attempt is a first delivery under a new id
unknown keeps its id
Reading `dispatchState` at all is a pre-existing defect, fixed here because
the false "sent" cannot be removed without it, and scoped to the send path.
`mutate`'s rotation for prompt/option/cancel plans is untouched. The
rejection copy is the desktop's notice, so an internal reason
(`provider_write_failed: ...`) still never reaches a person.
Tests: the hook test that was flipped to assert a rotated id now pins the
opposite -- one id across an ack-loss and two `unknown` replays, each
reported `unknown` rather than `accepted`. The send fixture grew the durable
submission row a real host returns; without it every send test asserted
against a shape that cannot express the bug.
* fix(native-chat): enforce fail-closed structured send replay
* fix(native-chat): align retry and mobile RPC contracts
* fix(native-chat): keep transient admissions retryable
* test(tab-bar): expand nested create menu in harness
---------
Co-authored-by: Merge Sim <sim@local>
|
||
|
Brennan BensonandMerge Sim
|
1ae7aa8bb4 |
feat(native-chat): resume an Agent Session History row into a new structured chat (#19176)
* feat(native-chat): resume an Agent Session History row into a new structured chat A Claude or Codex row in Agent Session History gains "Resume in New Chat": it opens a new structured native-chat tab that continues that provider conversation, with the prior turns already in the journal. Until now those rows could only be resumed into a PTY terminal; the structured branch could reveal a chat Orca already owned but could not adopt one it had never held. Almost all of the machinery existed. Both lanes already resume from the record's provider handle chain, the journal already has a transcript importer, and the handle chain already models `adopted` as an origin. The gap was that a create always minted an empty chain, so the adapters started a fresh conversation. This seeds that chain. The client names only the conversation. `agentSession.create` is reachable by paired mobile clients, so the transcript path and the account home are derived by the executing host and validated against the account homes it recognises — a client-supplied path would choose which file the host imports and which credential directory the provider child launches against. Failure refuses rather than degrades. A transcript that cannot be found refuses before anything is created; one that fails or decodes empty *after* the provider has resumed fails the attach, tearing the child down and publishing no tab, because an empty journal beside a context-carrying agent claims a continuity the provider never gave. Codex can resume into any workspace since it is handed the rollout path; Claude resolves transcripts under a project key derived from the launch cwd, so it is offered only for the workspace the conversation was recorded in. * fix(native-chat): widen adopted-home discovery and keep ordinary launches untouched Three corrections from review of the first commit. The adoption's account-home candidates now include the extra Codex homes session discovery already scans. A row this host listed could otherwise refuse to resume, which reads as the feature being broken rather than as a scope. Ordinary launches call `createStructuredAgentSessionLaunchIntent` with two arguments again. Passing the resume source unconditionally appended a trailing `undefined` that four existing call-site assertions had to absorb; the churn was the caller's fault, not the tests'. The transactional adoption guard's comment claimed the self-exemption is what lets a committed create replay. It is not: replay is settled earlier by the operation ledger, and an adoption always arrives with a null expected fence, so a request naming an existing session id is refused a few lines below either way. The exemption is part of what "another record" means, and the comment now says that instead. * fix: preserve history adoption through create and retries * fix: replay committed history adoption from durable identity * fix: validate history before claiming adopted sessions * fix: extract AI vault resume domains * fix: recognize typed history resume refusals --------- Co-authored-by: Merge Sim <sim@local> |
||
|
a65332a8bd |
feat(claude): move structured native chat onto the Claude Agent SDK and enable it on macOS and Linux (#18560)
* Join structured attach teardown through journal bind * fix: restore structured chat parity * feat: add Claude structured session adapter * fix: harden Claude structured adapter * fix: close Claude adapter edge cases * fix: start Claude init deadline after launch * feat: wire Claude structured sessions * fix: harden Claude structured runtime * fix: fence Claude structured compatibility * fix: preserve Claude free-text prompt answers * fix: decode addressed Claude prompt text * feat: enable Claude structured chat on mobile * fix(mobile): keep structured chat provider-aware * fix(mobile): negotiate Claude structured tabs * fix: keep scoped RPC tests native-free * fix: secure mobile structured image delivery * fix: close structured session data-loss gaps * fix: prove real Claude structured startup * fix: consume pre-spawn proof before retry * feat(native-chat): add desktop structured sessions * fix(native-chat): satisfy structured session cleanup gates * fix(native-chat): keep structured renders pure * fix(native-chat): open composer pickers upward * fix(native-chat): use existing view for structured sessions * fix: harden structured desktop status projection * fix: close structured desktop lifecycle gaps * fix: fence structured AI Vault resumes * fix: fence structured AI Vault resumes * fix: preserve structured tabs during activation * feat: toggle structured sessions between chat and TUI * fix: harden structured session handoffs * fix: bind structured TUI before rollout proof * fix: complete structured chat round trips * fix: align structured TUI return readiness * fix(native-chat): make reverse handoff transactional * Add Claude structured TUI handoff seams * fix(native-chat): clear sticky handoff recovery * fix(native-chat): complete mobile reverse after TUI exit * fix(native-chat): keep TUI transcripts readable * fix(native-chat): recover TUI transcript gaps * fix(native-chat): recover claimed TUI owners * fix(native-chat): retain cold TUI proof authority * fix(native-chat): preserve Claude handoff authority * fix(native-chat): recover TUI transcripts read-only * fix(native-chat): harden Claude handoff recovery * fix(native-chat): serialize structured handoff recovery * fix(native-chat): close handoff admission races * fix(native-chat): validate pinned launch environment * fix(native-chat): revalidate restored and retried owners * fix(native-chat): gate restart recovery publications * fix(i18n): catalog Claude session controls * fix(native-chat): wait for structured TUI process proof * fix(native-chat): queue stale idle TUI handoffs * fix(native-chat): route structured Codex options directly * fix(native-chat): persist structured session options * fix(native-chat): hydrate resumed structured options * fix(native-chat): preserve options across structured handoffs * fix(native-chat): replay pending option mutations * fix(native-chat): rotate settled handoff operations * fix(native-chat): rotate refused send operations * test(native-chat): derive refusal retry state from host * test(native-chat): give the host-oracle matrix test an explicit timeout * fix(native-chat): keep Claude option controls idle * fix mobile structured first-send hydration race * fix(native-chat): preserve handoff launch authority * fix(native-chat): harden shared handoff recovery * fix(native-chat): serialize structured handoff recovery * fix(native-chat): close handoff admission races * fix(native-chat): validate pinned launch environment * fix(native-chat): revalidate restored and retried owners * fix(native-chat): gate restart recovery publications * fix(i18n): catalog structured session recovery control * fix(native-chat): wait for structured TUI process proof * fix(native-chat): queue stale idle TUI handoffs * fix(native-chat): keep structured recovery provider-neutral * fix(native-chat): drop local terminal topology from structured sync * fix structured outbox and tab restore races * fix(native-chat): preserve Claude question groups * fix structured provider visibility and request handling * fix structured session TUI handoff recovery * fix reverse structured session handoff * fix(native-chat): recover Claude outbox and resume state * chore(mobile): preserve the working-tree lockfile state before the main merge Carries the pre-existing uncommitted mobile/pnpm-lock.yaml modification into history so the main merge cannot overwrite it. Verified benign pnpm drift (babel 7.29.7->7.29.8 transitives plus deprecation metadata); drops no patchedDependencies (the mobile lockfile declares none). * test(native-chat): drop orphaned Claude handoff-auth test left by the main merge 'pins Claude handoff auth through the terminal provider boundary' is absent from main and its production counterpart preserveClaudeAuthEnv no longer exists outside this test - orphaned residue of the terminal/native handoff work this PR excludes by scope. Removed rather than repaired: the failure was a renamed field (providerHome -> providerRoot), and renaming it would have carried out-of-scope handoff code into the merge. Body preserved as evidence and logged in CLAUDE-STRUCTURED-DISPOSITION-TABLE.md. * Fix mobile structured turn state * fix Claude structured session blockers * fix claude structured lane blockers * fix Claude acquisition exit proof * fix(claude): route stream-json launch through process wrapper * fix(claude): gate structured chat support * Fix Claude structured launch gating * fix(claude): split session acquisition and prune mobile scope * test(claude): align structured session fixtures * fix(agent-session): preserve handoff launch arguments * fix(claude): open journals through the factory after origin/main split The journal opener moved to journal-store-factory on main; retarget the Claude structured tests that still imported the old path. * fix(claude): resolve Claude structured launch args, auth, and win32 proof The origin/main merge re-expressed the lane's Claude wiring onto main's split orca-runtime facade and dropped three wires past green typecheck and lint. - resolveLaunchArgs discarded its provider parameter, so structured Claude sessions were launched with Codex app-server flags; Claude exits on --dangerously-bypass-approvals-and-sandbox, and a Codex arg-parse throw could block Claude session creation outright. - resolveClaudeLaunchEnv was no longer supplied, so the launch resolver fell back to the whole process env as configuredEnv and buildClaudeChildProcessEnv re-applied every auth var it had just stripped. The resolver now merges the Claude overlay onto a strip-applied copy of the inherited env, which also keeps PATH intact for withCliRuntimeOnPath. - The windowsProcessStartTimeAvailable producer was gone while the contract field and both consumers survived, so the renderer gate fail-closed and structured native chat was unreachable on every win32 host. Separately, structured Claude pinned CLAUDE_CONFIG_DIR unconditionally. An explicit pin makes the CLI abandon the macOS Keychain even when it names the CLI's own default, so a default claude.ai account could not authenticate where the legacy Claude terminal could. Pin only a home the CLI would not resolve on its own, matching ClaudeRuntimePathResolver, and compare against the env the child would otherwise inherit so a diverging overlay cannot outrank the record's account home. Also await the now-async revealNativeSession in its regression test, and set the native status before revealing so a rejecting reveal cannot leave a session released but never marked native. Claude-Session: https://claude.ai/code/session_013UqKCRB6k5e8UaYhXUHeWY * fix(claude): scrub case-insensitive Windows auth env * fix(native-chat): settle handoff outcome-write failures instead of leaking them A store write failure while recording a handoff outcome escaped the flow runner's catch handler, so the client never received the failure and the flow surfaced as an unhandled rejection (seen as an intermittent agent_session_store_corrupt error in the proven-dead-retry suite, whose teardown raced the flow's trailing outcome write). Record the failed outcome best-effort, and drain the coordinator before that test's teardown removes the store root. Claude-Session: https://claude.ai/code/session_011aXkcHyeiRJuezupQdjZaM * fix(native-chat): make the structured close-failure toast provider-neutral The structuredSessionCloseFailed toast fires for any structured session, but its copy said 'Codex chat', so a Claude structured session that fails to close showed the wrong provider name. The launch-failure toast is only reachable behind the agent === 'codex' gate, so its copy stays as is. Claude-Session: https://claude.ai/code/session_013ugSpCx4AWkySaJb69BQax * fix(native-chat): wire structured handoff proof recovery * fix(native-chat): wire structured handoff proof recovery * fix(native-chat): correct the structured chat opt-in copy The one `experimentalStructuredNativeChat` toggle gates both providers — `useStructuredAgentSessionCreate` runs `canUseStructuredNativeChat` for `'claude'` as well as `'codex'` — but its description named only Codex. Its scope line also said Windows keeps using terminal chat, while the gate refuses win32 only until the host proves it can read a process start time. `structured-native-chat-availability.test.ts` already pins that Windows is allowed once the proof is cached, so the two contradicted each other. Claude-Session: https://claude.ai/code/session_01RJFsidQWmKYFmeoUuVu4Tp * test(claude): pin @anthropic-ai/claude-agent-sdk 0.3.251 contracts against a scripted CLI PR 1 of the SDK migration: dependency + test-only harness, no product wiring. - Pin @anthropic-ai/claude-agent-sdk to exactly 0.3.251 — not the newest release — because 0.3.251 (published 2026-08-28) clears the repo's 3-day minimumReleaseAge supply-chain gate with no exclusion, while the newest release was minutes old and would have required excluding a brand-new publish from the exact control built to catch brand-new malicious publishes. Every contract this design depends on was verified identical on 0.3.251: the full option surface, no pid on SpawnedProcess (custom spawner stays mandatory), env defaulting to process.env when omitted, and --replay-user-messages appearing only via extraArgs. - Exclude all eight bundled CLI platform binaries via ignoredOptionalDependencies. The setting lives in pnpm-workspace.yaml because pnpm 12 no longer reads the package.json "pnpm" field (it warns and ignores it; verified by install ablation). Excluding the binaries is what makes Orca's pathToClaudeCodeExecutable override mandatory rather than merely preferred. Note: pnpm 12.0.0 honors the ignore list when reconciling an existing lockfile but not on fresh resolution of a new dependency, so the lockfile's SDK entry was pinned surgically; both 'pnpm install' and 'pnpm install --frozen-lockfile' verify clean and stable against the committed lockfile. - Contract-pin suite drives the real SDK against a scripted fake CLI and pins: unknown type/field/content-block pass-through (and keep_alive interception), spawner env fidelity plus the omitted-env process.env inheritance sharp edge, extraArgs producing --replay-user-messages, argument parity for every CLAUDE_STRUCTURED_BASE_ARGS entry plus --session-id/--resume/ --resume-session-at, canUseTool wire request_id stability and abort on control_cancel_request, one spawn per query, pathToClaudeCodeExecutable honored by the default spawner, the exact SDK version, and the eight platform binaries staying uninstalled. Claude-Session: https://claude.ai/code/session_01FGCRfYUnb4hbvfTAHGtJKQ * feat(claude): drive the structured transport through the agent SDK Replaces the hand-rolled `claude -p --input-format stream-json` transport with @anthropic-ai/claude-agent-sdk 0.3.251, keeping the existing connection interface for this commit so the acquisition path changes minimally. The control-plane rewrite is a separate change. Orca still supplies the process. `spawnClaudeCodeProcess` routes through `spawnProcess`, retains the child and its pid — the triple the durable lease adjudicates on — drains stderr so exit errors keep their tail, and hands `.cmd` shims to Orca's Windows argument encoder rather than the SDK's plain spawn. `close()` keeps Orca's own bounded tree-kill and exit deadline, so it still resolves true only after an observed exit. Launch resolution emits an SDK options object instead of argv; durable `launchArgs` translate to a typed option where one exists and to `extraArgs` otherwise, refusing a token neither can carry rather than dropping it. The child env is always passed explicitly — omitting it would let the SDK inherit `process.env` and reintroduce the ambient `ANTHROPIC_*` leak. The stdout line parser is deleted; the SDK owns framing, and unknown frames still reach the translator verbatim. Claude-Session: https://claude.ai/code/session_01JMhFjh9HEnkcJ5YTfCdgD3 * fix(claude): settle the frame the SDK pulled but never wrote The SDK's input pump is `for await (frame of prompt) { await transport.write(frame) }`. When that write rejects — the child dies between Orca's liveness guard and the write — the for-await ends abruptly and calls the generator's `return()`, so the code after `yield` never runs. The frame was already shift()ed out of `queued`, so the later `fail()` from the exit path could not reach it and `send()` never settled: `dispatchClaudeTurn` awaits that send before it can return `unknown`, wedging the caller and the durable outbox. The pre-SDK transport rejected on the stdin write callback instead. Retain the in-flight entry and settle it from the generator's cleanup, and let fail() reach it too for the pump that never resumes at all. Claude-Session: https://claude.ai/code/session_01AobxxokqQ3qcxS7sy7ckum * fix(claude): keep the agent SDK behind the structured-Claude boundary The ordinary OrcaRuntimeService graph statically reaches the Claude adapter and so the transport module, whose first line imported @anthropic-ai/claude-agent-sdk. The SDK is evaluated whenever the regular runtime loads, before any structured Claude session is chosen: it sets process.env.NoDefaultCurrentDirectoryInExePath, changing Windows executable resolution for later subprocesses, and a missing or incompatible install would break normal runtime startup — for a user who never leaves the terminal/TUI path. Defer the SDK to the connection, memoized so it loads once per process, and add the import-graph ratchet: a walk from the Electron main entry that fails on any static import of the package, plus a clean-fork check that loading the runtime leaves the Windows search variable untouched and a child-process pin that the side effect is still real. Claude-Session: https://claude.ai/code/session_01AobxxokqQ3qcxS7sy7ckum * fix(claude): answer list_models so the picker stops serving the seed sendControlRequest had no list_models case, so every request hit the default reject; readClaudeStructuredSessionOptions swallows that with .catch(() => null) and falls back to the static catalog. Every structured session therefore served a hardcoded model list with no per-model effort levels, no resolvedModel and no default detection, and nothing surfaced the failure. The pre-SDK transport got the live catalog from the CLI. Route it through the SDK's supportedModels(), wrapped in the { models } envelope the existing parser reads. Claude-Session: https://claude.ai/code/session_01AobxxokqQ3qcxS7sy7ckum * fix(claude): reap the child's descendants before killing it The forced step of the exit ladder went through the Codex helper, which spawns `pkill -KILL -P <pid>` and SIGKILLs the parent in the same tick: the parent usually dies first, the descendants reparent to pid 1, and `-P` matches nothing. An MCP or launcher descendant of a stubborn Claude child was left running. The test named for that requirement declined to assert it and killed the survivor by hand instead, so it could not fail for the thing it was named after. Route the Claude reap through Orca's existing sweep, which snapshots descendants while their parent link still exists and signals them before the root goes, and on Windows uses the identity-gated `taskkill /T /F`. The test now asserts the descendant is dead; the manual kill stays only as a failure-safe. close() still returns true only on an observed exit. Claude-Session: https://claude.ai/code/session_01AobxxokqQ3qcxS7sy7ckum * fix(native-chat): merge the duplicated handoff type import CI's static-analysis lint (`oxlint --config config/oxlint-code-quality-native-plugins.json src config tests mobile --deny-warnings`) exits 1 on the two separate `import type` statements from the same module. Claude-Session: https://claude.ai/code/session_01AobxxokqQ3qcxS7sy7ckum * fix(claude): answer a permission callback whose signal already aborted settleFrom registered the abort listener and then delivered the request. A callback that arrives already aborted never fires that event, so the promise stayed pending behind a durable prompt with no cancel path. Check the signal first, emit the cancel, and resolve the SDK's null sentinel without registering. Claude-Session: https://claude.ai/code/session_01AobxxokqQ3qcxS7sy7ckum * test(claude): wait for the child to record the frame, not just for its report The scripted CLI writes its report at startup, so `until(readReport)` returned a report with no user messages whenever the child had not yet read the line. The assertion then failed under parallel load. Poll for the frame instead of for the file. Claude-Session: https://claude.ai/code/session_01AobxxokqQ3qcxS7sy7ckum * fix(claude): coalesce partial deltas onto one assistant item and stop painting result frames Under --include-partial-messages every stream_event frame carries its own uuid, and the final assistant frame for a block carries yet another; only message.id ties them. The translator keyed each delta by its frame uuid, so a reply painted as one bubble per delta chunk followed by a complete duplicate under the final frame's uuid. The block's first stream frame now mints the claude:(sessionId, uuid) identity, deltas coalesce onto it through the shared 60ms seam, and the final frame reconciles onto that same item. Known SDK bookkeeping no longer reaches the provider-fallback row: result subtypes are catalogued and settled by the turn lifecycle, an empty thinking block (redacted thinking) is a modeled kind, a string-content user replay is a text block, and an empty user frame paints nothing. An unmodeled result subtype or content kind still lands on the bounded fallback row. Claude-Session: https://claude.ai/code/session_01GaP5HpYQbvy2hYehVhwfEW * fix(claude): prove descendant exit at the close boundary instead of on an unref'd timer close() reported proven=true as soon as the direct child exited while the descendant sweep's SIGKILL sat on an unref'd 2 s timer, so a SIGTERM-resistant MCP server outlived the lease release. The reaper now composes the same shared primitives the Codex structured provider uses: snapshot, verified bounded descendant termination on POSIX, taskkill /T /F on Windows. The proof is false whenever descendants outlive the deadline, a retried close re-verifies the retained snapshot rather than trusting the dead root, and the raw pipe child no longer goes through the PTY job sweep it never owned a job for. Measured on macOS: a killed child of a SIGSTOPped parent stays a matching zombie row in ps, so the root is killed while verification runs rather than stopped first as the Codex non-group path does. Claude-Session: https://claude.ai/code/session_0161QFm3KVRNJKfdzWVGVNWk * feat(claude): replace the hand-rolled control plane with the SDK's native surface PR 3 of the Claude structured SDK migration removes the wire-frame scaffolding PR 2 kept, so Orca drives the SDK's typed control surface directly. Inbound permissions move from a rebuilt control_request dispatch to the SDK's canUseTool / onUserDialog callbacks. The prompt registry now carries the callback's own resolver: a decodable can_use_tool becomes a durable prompt whose answer settles the callback; a malformed one is denied without registering; the SDK's abort signal (fired on control_cancel_request, which the SDK matches and dedups itself) forgets the prompt and settles it null, and a late answer after abort finds no prompt and is refused. Closing settles every in-flight callback so no promise dangles. The claude-agent-sdk-control-bridge that rebuilt the wire frame is deleted. Outbound control maps to Query methods: interrupt() for cancel, setModel / setPermissionMode / applyFlagSettings for options, supportedModels for the model list, initializationResult() for init proof, each under Orca's own request deadline and error classification. Cancel is interrupt-receipt aware: a CLI advertising interrupt_cancel_queued_v1 gets cancel_queued in one round trip, otherwise the receipt's still_queued uuids are swept with cancel_async_message so a cancelled turn cannot spawn a later unexpected turn; older CLIs resolve no receipt. Init keeps the 10s deadline and the unauthenticated-startup guidance. Every behavior is failing-first and ablation-proven; the toggle-off import boundary and the accepted loss of unknown-control visibility rows are unchanged. Claude-Session: https://claude.ai/code/session_01Pqjduxt5G4rr9aYvtp7rNm * fix(claude): arm the descendant snapshot before stdin closes and make the tree verdict unproven by default A healthy Claude root leaves within the graceful window, and the close ladder only snapshotted descendants when the root was still alive after that window. So the common close never looked at the tree: `treeExited` stayed null, `!== false` passed it, and close() reported a proven exit with an MCP child still running. A root that died before the walk made the snapshot vacuous too. The proof is now unproven by default. The reaper holds one verdict in Orca's vocabulary (exited / live / unverifiable), assigned in exactly one place from the bounded verification, and close() returns true only on `exited`. The snapshot is armed before stdin closes, while the root can still be walked, and is verified after the root exits; a root that left before any snapshot could be armed stays unverifiable rather than vouching for descendants it never showed us. The shared verifier gains the three-way verdict behind its boolean face, and the connection reports the root and tree verdicts separately along with the child's exit status. One verification per close attempt: the retried close re-verifies, so the intra-attempt re-reap is gone from the teardown budget. Claude-Session: https://claude.ai/code/session_01BSmXgkWSsNHft8jFkdBFG9 * fix(claude): verify the Windows tree after taskkill instead of trusting that it ran `terminateWindowsProcessTree` resolves from taskkill's callback whatever the error says, so a timeout, an access denial, a recycled root and a surviving descendant all looked identical to the reaper — which then returned a proven exit unconditionally. close() reported true and the lease was released with an MCP descendant potentially still live. The Windows branch now snapshots the root's descendants while it is alive and, after taskkill, polls a fresh process table to a bounded deadline: a row still matching by pid AND creation time is `live`, an unreadable table is `unverifiable`, and only a table with no match is `exited`. Creation time is the PID-reuse guard the POSIX path gets from ps lstart, so a descendant that denied a creation-time query is omitted rather than signalled on a bare pid. A root already observed exited is never taskkilled: `/T /F` on a recycled pid would take an unrelated tree down with it. The captured tree is tagged by platform so neither verifier can be handed the other's rows. Claude-Session: https://claude.ai/code/session_01BSmXgkWSsNHft8jFkdBFG9 * fix(claude): release a reservation on a first-hand root exit instead of latching it into manual recovery Making close() strict about the descendant tree exposed a second defect at the same boundary. A create-time acquisition has no ownerProcess until publication, so an unproven cleanup mapped to handoffStage `manual-recovery`, and adjudication then refuses every later attach with agent_session_ownership_unknown. A user who was merely signed out, or whose --resume the CLI rejected, wedged the session id permanently. Each question now answers from its own evidence. close() is unchanged and stays strict about the tree. Separately, the lease is keyed on the root's pid and start time, so when Orca's own child handle observed that root exit and no descendant snapshot was ever admissible, the reservation is released and the CLI's exit code and stderr reach the user. A descendant observed still alive, or a root Orca never saw leave, stays unproven and keeps the reservation. The settlement records only what was observed: the released lease says the provider process exited and its descendants were not verifiable, rather than reusing the wording that claims cleanup proved no child remains. Claude-Session: https://claude.ai/code/session_01BSmXgkWSsNHft8jFkdBFG9 * fix(claude): surface an API error a result frame reports instead of settling the turn on it The SDK models an API failure as a SUCCESS-subtype result whose `result` string is the user-facing error text, with no assistant frame behind it. The translator suppressed every catalogued result subtype as turn bookkeeping, so that turn tombstoned its lifecycle and showed the user a completed, empty reply with no sign anything had failed. Suppression is now by meaning. A result reporting a failure routes to the bounded provider-error surface, leading with the provider's own sentence and keeping the raw frame behind the row's disclosure; ordinary successful results stay off the timeline as before. A turn the user aborted also stays suppressed: its interrupt frame already says so, and its execution diagnostic would only be noise on every stop. Claude-Session: https://claude.ai/code/session_01BSmXgkWSsNHft8jFkdBFG9 * fix(claude): drop the stream state of turns that never received their final frame Every streamed delta recorded its block's identity, latest text and checkpoint length. Only the final assistant frame removed them, so an interrupted turn left its whole accumulated reply reachable until the session was disposed, and a long session with repeated interruptions grew those maps without bound. The partial text was already journaled by the flush that precedes settlement, so the live copy was pure retention. That state now lives in its own module, named for what it does — grow a streamed block's journal row between its deltas and its final frame — and turn settlement drops every block still awaiting a final. The translator reports how many remain, which is the invariant: a settled turn leaves none. Also makes a timed-out process-table read retryable while the root is still alive. A loaded host can miss the table's one-second deadline, and latching that as "no descendants" both lost the descendant sweep and, on a busy machine, made the close ladder report unproven for a tree it never actually looked at. Only the root's death still makes a missing snapshot final. Claude-Session: https://claude.ai/code/session_01BSmXgkWSsNHft8jFkdBFG9 * perf(claude): capture the Windows descendant tree from one process-table read The capture walked the descendant tree and then read the table again for the creation times the walk's projection drops. Each read is bounded in seconds and both run inside the close ladder's budget, so the second one cost the worst-case teardown three seconds for data the first read already held. The walk is now exported from the module that owns it and runs over rows the caller has already read, which is also what lets the snapshot keep the PID-reuse guard the projection cannot carry. Claude-Session: https://claude.ai/code/session_01BSmXgkWSsNHft8jFkdBFG9 * fix(pty): spend the descendant verification window instead of surrendering on one slow table read The verification abandoned the whole check the first time a process-table read missed its own one-second deadline, with seconds of its window still unspent. On a loaded host that reported a tree unverifiable without ever having looked at it, which the Claude close ladder then turned into an unproven close and a retried teardown. It also made the descendant-exit tests flake under a parallel suite run, for the same reason and with the same honest-but-premature verdict. A read that missed its deadline is now simply not an answer: the loop waits and reads again until its own deadline, and only a window that ends without a readable table reports unverifiable. This can only turn a premature verdict into one backed by evidence; it never manufactures a proof. Claude-Session: https://claude.ai/code/session_01BSmXgkWSsNHft8jFkdBFG9 * fix(claude): never let a later failed look collapse an observed live descendant into unverifiable The reaper's single assignment site latched only 'exited', so a second reap whose table reads all missed their deadline overwrote an earlier completed verification's 'live' with 'unverifiable'. The acquisition release gate discriminates on exactly that pair, so a root exit after such a decay released the lease over a descendant that had been observed alive. The latch is now monotone in trust order: exited is final, and live is only ever raised to exited. Claude-Session: https://claude.ai/code/session_01HfdhsvSJucLw4cTZxzg2CP * fix(claude): never prove a Windows tree gone while a descendant denied identification The Windows snapshot dropped rows that denied the creation-time query, and an emptied snapshot was judged exited without any table read: a descendant Orca was refused information about was treated as one that had left. The snapshot now counts the unidentified rows it saw, and verification caps its verdict at unverifiable while any exist. Nothing is ever signalled on a bare pid, as before. Claude-Session: https://claude.ai/code/session_01HfdhsvSJucLw4cTZxzg2CP * fix(claude): classify cleanup after a first-hand exit as a root exit instead of a proven tree When the CLI died between a successful acquire and the host's commit or proof of the lease, handleExit had already removed the session, so releaseAcquisition found nothing and reported true. The attach flow then settled exit-proven with deathEvidence claiming cleanup proved no provider child remains, though the tree was never verified. The adapter now keeps the exit that removed a published session until the session is acquired again; acquisition cleanup runs that connection's close ladder and classifies its verdict exactly as a start-time failure would be, so the record reads root-exit-observed. The wire helper keeps that typed classification and its provider diagnostic instead of wrapping it as unproven, and the router gives up its owner even when the release throws. Claude-Session: https://claude.ai/code/session_01HfdhsvSJucLw4cTZxzg2CP * fix(claude): integrate SDK teardown and picker lifecycle fixes * fix(claude): preserve resume leaf and settle processless spawns * fix(claude): reacquire from persisted resume leaf * fix(native-chat): restore Claude grouped question handling * fix(claude): persist only resumable transcript leaves * fix(claude): recover structured session exits safely * fix(claude): close remaining structured session P1s * fix(claude): harden transcript branch proof * Remove superseded root fix reports * fix(windows): restore indexed descendant row walk * fix(router): forward force-close lifecycle * fix(claude): fence stale turn cancellations * fix(claude): fence cancellation after unknown dispatch * fix(claude): fence replay and option recovery races * fix(claude): block replay fallback after waiter eviction * fix(claude): fence evicted slash results * fix(claude): fence ambiguous results and restore options safely * fix(claude): scrub SDK child env and localize pending launch * fix(claude): pin transcript roots and exit recovery proofs * fix(claude): retain unproven SDK exits * fix(claude): settle retained exit before reacquire * fix(claude): resume from settled retained cursor * chore: remove tracked review artifact * fix: harden Claude SDK transport session cleanup * fix: close Claude sessions safely * fix(claude): close races with fresh child snapshots * fix(claude): fail closed on recycled child identities * fix(claude): gate root cleanup on process identity * fix(claude): fence same-second root identity reuse * fix(claude): restore the root SIGKILL fallback the identity gate took away The direct root kill goes through the handle Node owns, not through a pid: libuv drops that handle in the same turn it reaps, so the signal either reaches the process Orca spawned or reaches nothing at all. Gating it on a process-table probe therefore bought no safety and cost the tree its only fallback whenever the probe declined -- a first capture landing in the fork's own second, a recycled descendant pid voiding the snapshot, or a process table that could not be read on either platform. Identity verification stays where a bare pid is genuinely addressed: Windows `taskkill /T /F`, and the descendant sweep's own revalidation before it signals. Also stops a declined root probe from collapsing an observed `live` or `exited` descendant verdict into `unverifiable`, and stops a successful taskkill from reporting `unverifiable` because a later probe found the root correctly dead. * docs(claude): rewrap the root-kill ordering comment * Match the Claude structured launch to the terminal path's managed-account auth rules The SDK path stripped ambient Anthropic auth unconditionally, let an explicit agentDefaultEnv override beat a pinned managed account, and had no account-switch guard. Reuse the terminal preflight's own predicate and messages so both transports strip, refuse, and report identically, and cover the CLI transcript location that mobile native chat depends on. * Reach the Claude structured chat lane from the desktop UI The main process has had a complete, correctly gated Claude Agent SDK lane for a while, but no renderer ever asked for it: the launch route accepted only `codex`, and the create path was typed `agent: 'codex'` end to end. Widen both to the structured provider union that already exists (`AgentSessionHandleProvider`), and generalize the codex-named create path instead of adding a Claude twin beside it. The pending-launch registry is now keyed by agent as well as workspace — a shared key handed a second caller the first agent's intent, so a Claude and a Codex launch in one worktree collided. Windows, per agent. Codex's client-side win32 refusal is deliberate and settled elsewhere, so it stays exactly as it was. Claude's answer is no longer guessed from the client's platform: a structured session fences its provider child on that child's process start time, and only the executing host knows whether it can read one. `agentSession.createSupport` already answers precisely that, per agent, and had no renderer caller — so the Claude create path asks it before creating and turns a "no", or a probe it cannot get answered, into the definitive refusal the launch fallback already handles. Fail closed either way. That refusal mapping also closes a real gap: the host reports an unsupported location by throwing `structured_agent_session_unsupported`, which reaches the client as a transport rejection rather than a refusal envelope, so `StructuredAgentSessionCreateRefusalError` never fired. The launch would retry the create, strand itself in `visibilityUnknown`, run no legacy fallback, and show an error toast. Close a fail-open hole while Claude and win32 become reachable: `create` with a client-supplied location, and `ensure`, both skip the worktree-resolving support check. They now ask the executing host the same question directly, so a host that cannot fence a provider child no longer creates one on a client's say-so. Also deletes `structured-agent-session-provider-routing.ts`, a duplicate of `structured-agent-session-provider-support.ts` with no importers. WSL, SSH and paired hosts, floating workspaces, draft prompt delivery, explicit TUI customization and initial session options all keep refusing; folder workspaces keep working. * P1-1: make the structured Claude auth policy required and testable The optional dep plus a {stripAuthEnv:false} fallback meant a dropped wiring under-stripped silently. Required at all three hops, asserted at install time for the @ts-nocheck caller, and the settings-to-policy mapping is now a named tested function. * P2-3: mobile's default Claude transcript root must follow CLAUDE_CONFIG_DIR session-file-resolver's default ignored the variable the pinned account home follows, so a CLAUDE_CONFIG_DIR launch wrote one tree and mobile read another. The Task-4 test now resolves with no root override (mobile's own call) and checks the answer against the root the CLI itself reports, instead of mirroring the code under test's own expression. * P2-1/P2-2/P3: close the teardown window, join the live-auth gate, align the refusal P2-1: a switch beginning inside the acquire teardown left a dead chat and no replacement. Past that point the launch waits the swap out and refuses only if it never settles; the entry guard still refuses outright, because nothing is torn down there yet. P2-2: structured children now hold the same OAuth-refresh gate a Claude PTY does, so a managed refresh cannot rotate the token out from under a live turn. P3: the refusal now matches the strip it guards (case-folded on win32, presence not truthiness), and the dead structured-to-TUI builder states its auth policy instead of silently signing a system-auth user out. * Make the live-auth gate tests independent of sibling connection teardown order * Do not offer structured Claude under a WSL-only managed account Structured Claude launches against the ambient Claude config, which the account service keeps in sync with the selected HOST account. A WSL-bound managed account lives inside the distro and is never synced there, so on Windows a structured session would authenticate as whatever the ambient identity happens to be while the UI names the WSL account — the user is told one identity and given another. That was unreachable only because nothing offered structured Claude on win32. Enabling it makes it reachable, so gate it here rather than patching the auth layer: refuse the structured path when the active managed Claude account is WSL-bound, and let the terminal-backed path — which resolves the account per runtime — handle that account shape. The answer rides the agentSession.createSupport seam the renderer already consumes, so no new capability and no renderer knowledge of account internals. A create the host declines becomes the definitive refusal the launch fallback already turns into a legacy native chat tab, with no error toast. Unknown answers refuse. An install with no managed accounts claims no identity and is fine, but an active selection that cannot be resolved — or account state that cannot be read at all — is not evidence that the ambient identity is right. Claude only. Codex resolves its account through a different path and its createSupport answer is untouched, as is every Codex routing decision. * Read the structured Claude account gate through the auth policy's accessor The gate resolved the active account from the account-service snapshot's runtime map; the auth policy resolves it with getSelectedClaudeAccountIdForTarget(settings, { runtime: 'host' }). Those are two sources and two resolution rules, and they disagree on a legacy settings blob that carries the selection only in the flat activeClaudeManagedAccountId: the accessor falls through to it, a direct read of the runtime map does not. The gate would then refuse a launch the policy would have run under host-1 — and in the mirror case a session could be admitted under a policy computed from a different account than the gate approved. Read the same settings through the same accessor so agreement is structural rather than coincidental, and drop the controller accessor that existed only to reach the snapshot. No behaviour change for any state both already agreed on; Codex is untouched. * Round-3 review fixes: N-1 empty-value regression, N-2 gate leak window, N-4 lost history N-1: my presence-based conflict predicate refused a terminal launch that works today. 'ANTHROPIC_API_KEY=' is how a user blanks a variable and the settings pipeline preserves that empty value; an empty override cannot beat the pinned account and the strip removes the name anyway. Back to truthiness for the value, keeping the win32 case folding. N-2: enter the live-auth gate only after the exit/close handlers that release it, so no throw in between can leave an entry nothing reconciles. N-4: the Claude transcript resolver searches config-dir-then-default and de-dupes, matching the Codex sibling in the same file, so adopting CLAUDE_CONFIG_DIR no longer hides history written before it. * Run the managed-account gate on every Claude acquisition, not just create createSupport gates the create path, but a session's account state can change while it lives. A reacquire after an unexpected child exit re-resolves the launch and re-derives auth, with nothing re-checking the gate — so a session created while supported could come back up in the refused shape. With the strip predicate keyed on there being an active non-WSL account, the WSL-only user's normalized steady state (accounts exist, none active) does not strip, and that reacquire reaches the child with ambient auth while the UI names the account. Gate at resolveLaunch, the one choke point every acquisition passes through, refusing with the pre-spawn error the caller already handles. Same predicate as create-time, now sharing one settings reader so the two cannot drift. Claude only; Codex resolves its account on a different path and is untouched. The runtime class that wires this does not typecheck its own `this` calls — a missing hookup compiles clean — so the wiring is pinned behaviourally rather than trusted to the compiler. * Move the structured Claude gate out of the @ts-nocheck runtime files Both call sites of the managed-account gate sat in files whose first line is `// @ts-nocheck`, so neither was typechecked: three arguments to a one-argument function plus an undeclared identifier compiled clean. New auth-identity decision logic had no compiler behind it. Move the verdict into a checked module that takes the two facts the runtime owns — the adapter's answer and a settings getter — and decides. The runtime class now only forwards. Move the gate reader's construction into the checked installer too, so the nocheck file passes a plain settings closure and never names a gate symbol. Every reference to the gate predicate and its reader now lives in a checked file, so the ablation that used to pass silently is a compile error at both the create-support and reacquire sites. Removing the file-level @ts-nocheck is a separate, larger job and is not attempted here. * Derive the gate test's auth policy from the settings under test A hardcoded stripAuthEnv asserts a gate/policy pairing production cannot produce, and false additionally lets launch.env inherit the runner's real process.env. Derive via claudeStructuredAuthPolicyForSettings instead: the gate settings type is the same Pick the policy takes, and both resolve the account through getSelectedClaudeAccountIdForTarget. * Pin the absent-vs-empty distinction in the managed-account gate An empty claudeManagedAccounts array is a real answer: the user has no managed accounts, nothing claims an identity, and the ambient path is legitimate. A readable settings object with no such field is settings we failed to parse — the same unknown as unreadable — so it refuses. The two are one character apart in the code and the difference is invisible without the reasoning, so record it at the branch and pin both sides. The test fails under the obvious "consistency fix" of treating a missing field as empty. * fix(claude): keep command queue bookkeeping out of the transcript Claude Code 2.1.258 emits a `command_lifecycle` frame for every uuid-stamped command it starts, completes or cancels. The frame carries a command uuid and a state and no content, and the CLI keeps it out of its own transcript -- but it is absent from the SDK's SDKMessage union and so from Orca's frame catalogue, where an uncatalogued kind defaults to a substantive row. Every structured turn therefore painted raw JSON rows into the user-visible transcript. Catalogue it and disposition it as status chrome. The unknown-kind default stays `timeline-substantive`: a kind we have never seen is likelier to carry content than to be chrome, and a visible row we can catalogue later beats content we silently dropped. A lifecycle state that reads as a failure still surfaces, because the payload error check in `classifyProviderFrame` outranks the catalogue. * fix(claude): let a re-walked descendant become eligible for the forced sweep A descendant first observed by a capture inside its own birth second could never be SIGKILLed: `ps lstart` is second-resolution, so that capture cannot rule out a pid recycled later in the same second, and the merge pinned each retained row to the boundary of the walk that first saw it. SIGTERM-resistant children forked in that window were signalled and then never escalated -- they survived close, quit and restart, reparented to init, and had to be killed by hand. Advancing that boundary on any later capture would be unsound: a later capture matching pid, pgid and start-second is exactly what an impostor would also show. But a capture is not a match -- it is a fresh ppid walk from a root Node pins through its own handle, so a row it re-derives is proved ours at that instant without appealing to its start time. Chain the fence from there instead, and take that walk at the close boundary while the root certainly still lives: the root may leave inside the grace window, and the post-timeout refresh never runs. A row absent from the later walk still keeps its earlier boundary, and a row no walk has ever re-derived in a later second is still never escalated. * Treat an absent managed-account list as empty, not as unreadable An empty claudeManagedAccounts array and a missing one are the same answer: this user has no managed Claude accounts, so nothing claims an identity and ambient auth is the truth. Refusing on absence strands any profile that simply never wrote the key, and it disagrees with the auth policy, whose own predicate takes `(accounts ?? [])` for exactly this reason. Only settings that cannot be READ stay unknown, and those still refuse — as do a WSL-bound active account and a selection naming an account the list does not explain. The earlier reasoning treated a missing field as settings we failed to parse. That conflated "not present" with "not readable"; only the second is unknown. * Support structured Claude when accounts are registered but none is selected Registered-but-deselected Claude accounts were refused, which is behaviourally identical to having no accounts at all: the auth policy does not strip, ambient auth is the truth, and the UI names no host identity. A user who deselected their accounts silently got legacy chat with nothing explaining why. Nothing selected for the host runtime is two states the settings cannot tell apart after the fact, because pruneInvalidClaudeRuntimeSelection empties the host slot and persists null in the second one: honest deselection -> ambient auth, UI names nothing -> SUPPORTED the WSL-only steady state -> ambient auth, UI names the WSL account -> REFUSED The presence of any WSL-bound account in the list decides. Simplifying this to "none active -> supported" re-opens the auth-identity misrepresentation, so the tests fail loudly on exactly that: five of them, across the unit rule and the createSupport path. * Stop treating an unanswerable create-support probe as a refusal A worktree is not resolvable for a beat after createWorktree resolves, so a probe fired immediately after creation fails the RPC with selector_not_found instead of answering. The catch collapsed that into `supported = false`, so the composer refused and quietly built a terminal session — the gate never said no, it was never asked successfully. Elapsed time was the only input that decided whether a Claude launch went structured. "Could not answer" and "answered no" are different states and only the second is a verdict. Retry while the host cannot yet resolve the selector, with a bounded backoff that covers the measured window with margin, and keep refusing on the first ask for everything else. Fail-closed is unchanged: a probe that still cannot be answered when the budget is spent refuses. The retry is narrowed with the shared error-code matcher, which classifies a token that transports re-wrap into a longer message without matching prose that merely mentions it. Codex never probes, so this race has never been able to refuse a Codex launch — the race itself is identical for it. Recorded at the early return, because whoever gives Codex a probe inherits the bug. * fix(claude): fence the forced sweep on re-derivation, not on lstart's second A descendant forked in the same wall-clock second as every walk that sees it was signalled with SIGTERM and then never escalated, so a SIGTERM-resistant child survived tab close, app quit and a full relaunch. Two children of one parent 96ms apart across a second boundary took opposite paths. The leak predates this branch: it reproduces with the change reverted. `ps lstart` has one-second resolution, so a walk landing inside a row's birth second can never rule out a pid recycled later in that same second. But a walk is not a match: a ppid walk only reaches what the root actually parents, and the root is pinned by Node's own handle, so a row the walk re-derived is ours whatever second it was born in -- a stranger would have to have been forked into our tree, and then it is not a stranger. Fence the escalation on that. Rows a merge retained from an earlier walk are not re-derived and still answer to the start-time fence, which remains correct for them. Scoped to callers that revalidate identity before signalling, which is the Claude close path. Codex teardown reaches this same verifier and is unchanged; the argument holds there too, but widening it is its own deliberate change. Also reverts two changes from the previous attempt at this leak. Advancing the capture boundary on a later walk is inert once the sweep fences on re-derivation -- both key on the same set of rows, so the new term short-circuits for exactly the rows whose boundary it advanced. The extra ladder refresh was a duplicate full process-table read: close() already awaits tree.refresh() immediately before proveClaudeChildExit, on the only path that reaches it. Known property: the kill lands roughly a grace window after the walk that proved membership, so a pid recycled inside that gap could in principle be signalled. It is bounded -- matchingSnapshotRows already requires the live row to carry the same start-second and pgid, so an impostor must be born in the remainder of that one second, land on that exact pid, and sit in the same process group, and it has already received the unfenced SIGTERM from the same loop. * Run the Claude structured integration suite as a runtime client The suite exercises agentSession.* for Claude, not the mobile surface: nothing in it asserts anything mobile-specific and its sibling integration suites use 'runtime'. Mobile now additionally requires the experimental structured-chat setting, which structured-agent-session.test.ts pins in both states, so the stale 'mobile' fixture was claiming coverage it never had. * fix(claude): report effort from get_settings, which is the only frame that has it The composer's Effort pill rendered blank in every structured session. This is not a missing source: the publication reads `effortLevel` off the `system/init` frame, and that frame has never carried an effort of any kind, while the correct value is already fetched at acquisition and thrown away on the auth diagnostic. Verified two ways -- a live get_settings probe against Claude Code 2.1.258, and the shipped binary's own init frame construction, which lists `model` and no effort. So `reportedOptions.effort` was always empty, the options reader dropped the key, and the pill had no value. Model survived only because `currentModelId()` has a fallback chain. The get_settings call acquisition already makes reports the session's current effort as `effective.effortLevel`; pass that into the publication instead. Selecting an effort already worked, so this is the arrival value only. The legacy PTY path is unaffected and must not be "fixed" to match: it reads its effort by parsing the startup banner (`CLAUDE_MODEL_EFFORT` in src/renderer/src/components/native-chat/claude-terminal-session-options.ts), which is why it shows a value where the structured path does not. Also removes the fixture that hid this: the fake init frame invented `effortLevel: 'high'`, a field the CLI does not send, which is why every gate stayed green over a value that is always empty in production. The fixture's get_settings now returns the real {applied, effective, sources} shape instead of a bare `{env: {}}`, so the two adapter tests that asserted an effort keep asserting it through the path production actually uses. The reader returns null rather than defaulting: an effort nothing measured would repeat the fixture's mistake, and a blank pill is the honest degradation if the provider ever renames the key. * fix(claude): only record an effort the child confirms it adopted apply_flag_settings answers `success` for an effort it then ignores. Measured against Claude Code 2.1.258: applying `bogus-effort-xyz` returns subtype "success" with no error while `applied.effort` stays at its previous value, and a valid `low` moves it. The option write treated the absence of a throw as adoption and recorded the requested value unconditionally, so Orca would show and persist an effort the child was not using, with nothing anywhere reporting a problem. Read the effort back after applying it, through the same reader the arrival value uses, and reject when the child reports a different one. A readback that could not be taken is not evidence of a refusal -- the apply itself succeeded -- so it still records; only a readback that disagrees rejects. Not reachable from today's picker, which offers catalog values only, but the CLI's effort catalog is server-delivered and has changed before, so a retired id would otherwise become a pill confidently displaying a setting that never took. * test(claude): assert the effort contract against the real binary The blank pill survived every gate because the only tests that touched it were fixture-backed, and the fixture invented the field. A test that pins the shape we read cannot catch the provider renaming the key, which is the failure mode that produced this defect. Asserts both halves against a live authenticated CLI: that no frame it publishes carries an effort at all, and that the session's current effort arrives through get_settings. Which frame proves the session varies by host -- this machine proves it with a SessionStart hook rather than a system/init frame -- so the negative half asserts over every published frame rather than picking one. Skips with the rest of the file when no authenticated CLI is present. * fix(claude): stop the synthesised content-part kinds leaking into the transcript Sending an image put a bare `claude · message:user:content:image` row between the user's bubble and the answer. Two causes, and only the second is a family. An image part counted as modelled only when `source.type === 'url'`, but claudeDispatchMessageContent sends a local attachment as a base64 source and the CLI replays that shape back, so every attached image was classified unmodelled. Accept the base64 and file sources Orca itself sends. The family is the real defect. `message:<role>:content:<type>` kinds are synthesised at runtime from whatever `part.type` arrives, so unlike the top-level frame catalogue they can never be enumerated ahead of time -- the `?? 'timeline-substantive'` default then prints the synthesised name at a user who cannot act on it. That default is right for top-level frames, where "substantive" means show the frame; here it meant show our own vocabulary, which drops the content AND leaks the opcode. So an unrenderable part now renders a sentence saying exactly that, with the kind and payload still on the row's disclosure. A part that carries its own readable sentence keeps it -- the placeholder is a fallback, not an override. An unknown future part type is therefore visible, never silently dropped and never printed as a kind: the same principle as the effort readback, which records only what the provider confirms. * Declare agentSession.requestHandoff on the cross-version wire surface The manifest is a ratchet for cross-version reachability, so the method is declared with real HandoffParams rather than counted. requestHandoff is capability-gated through requireStructuredHost and has no client caller, so declaring it is the whole of the change. Also model two host capabilities the harness omitted: the stub host's supportsCreate, and the fake adapter's, without which adapterSupportsCreate falls through to a supportsLocation the fake also lacks. Every ensure was refused for the harness's silence rather than for its location. * Gate structured Claude session tabs on the client capability that names them The Claude structured lane deleted the projection's `agent !== 'codex'` filter and added CLAUDE_STRUCTURED_AGENT_SESSION_RUNTIME_CAPABILITY in the same commit, but never wired the constant to anything. Paired clients then received agent-session tabs for Claude, which no shipped client renders -- mobile's resolveMobileNativeChat returns null for every agent but codex, so the row listed and selected into a pane with neither chat nor terminal. Restore the filter behind the declared capability instead of the bare agent name. No client advertises it yet, so this matches main's behaviour today and becomes a negotiation a future client can opt into. * Confirm the structured Claude model against the model the CLI reports set_model answers success for any string, including a model it cannot resolve — the failure only surfaces when the turn runs — and get_settings reports the settings-file model, not the session's. The init frame that opens each turn is the only channel carrying the adopted model, so keep the session's reported model current from it instead of reading it once at acquisition. Also stop rejecting an effort the readback cannot represent: max is session-scoped and excluded from the persisted effortLevel, so a readback reporting the level underneath it is an absence of evidence, not a refusal. * Clear the session-option hedge when the provider confirms the value The pill claimed every option was unconfirmed for the life of the session: the renderer recorded each write as dispatched and nothing ever moved it, so a model the CLI had already reported back still read as unconfirmed. Carry the provider's own confirmation to the surface. Main reports which option ids the provider named rather than merely accepted, and the client re-reads options as a turn changes, because the frame that opens a turn is where the adopted model arrives. A value the provider has not reported stays hedged, including an effort whose readback could not be taken. The confirmed list is optional on the wire: a host that predates it sends nothing and the client keeps hedging, which is the behaviour it had. * Keep the model report current across an acquisition fence bump * Show the picked session-option value and let the provider report correct it The pill showed a "not confirmed" second tooltip line for any value we had sent but not yet seen reported back. Nothing acts on it, and for the PTY lane it was permanent — that transport has no report channel. The pill now shows the picked value immediately and the provider's per-turn report corrects it when the two disagree; a newer local write still outranks a report that precedes it. `dispatched` stays as a provenance member rather than collapsing into `applied`: it is produced independently by the PTY lane, and it is where the `confirmed` wire field lands, which would otherwise be unobservable. Effort keeps its readback and its rejection path. That matters more now, not less: with the hedge gone the rejection is the only user-visible failure signal on this surface, so a spurious one would be the loudest bug here. Skipping the readback for an effort the settings response structurally cannot echo is what prevents it — the response carries the persisted level, so reading it back for a session-scoped value would report the level underneath and fail a valid write. * Hedge a session-option value only when the terminal transport sent it Both lanes emit `dispatched`, so it could never say which one produced a value. The descriptor now carries the transport that built it, set once in the shared snapshot builder from a parameter that is required rather than defaulted — the builder is the only place a descriptor is constructed, so a new producer has to name its lane or fail to compile. The structured lane confirms every value from the provider's own per-turn report, which makes the hedge transient noise there. The terminal lane can only learn an outcome by parsing the screen back, and only for Claude: every other agent's `dispatched` value stays unconfirmed for the life of the session, so the line is the only signal that we sent something we never saw land. * Refuse an effort the session's model advertises no control for * Refuse tab mutations on a Claude row the client never negotiated The branch added a case asserting a client advertising only agent-session.structured.v1 may mutate a claude row. That is the same ungated behaviour the projection gate removes, encoded a second time — mutation authorization reads the projection, so hiding the row refuses the write. Assert that contract instead, and add the positive case for a client that does negotiate Claude rows. * Resolve the Claude session's current model in one place so the effort guard and the pill agree * Record an effort the child did not adopt instead of refusing the write apply_flag_settings answers success for an effort it then ignores, so the readback exists to detect that. Refusing on it made the detection a veto, and a veto is only correct if the readback can never be wrong about which model is current -- which it was, twice. The pre-flight guard already refuses a level the model advertises no control for, so the veto guarded a door that is now locked upstream. Keep the detection, drop the refusal: a disagreement records the child's own answer and omits the option from confirmed, so main stops vouching for a value the provider rejected without blocking the user's write. * Stop a slow whole-machine ps from being read as an absent process `ps -axo ...command=` pays a per-pid argv read: measured 1.15s for 1,948 processes (0.03s without `command=`), and CPU contention stretched the same capture to 6.0s. Two budgets sized for a cheap look then misreport a readable machine. The reader's 3s ceiling killed 6 of 20 consecutive captures at load 27, so every consumer answered "unverifiable" about a table it could read. Raise it to 15s, and stamp the capture instant at ps START so `capturedAgeMs` is the upper bound its contract promises -- a 6s capture used to report itself as freshly taken, understating staleness against a 5s kill gate. The TTL keys on completion so a slow capture still coalesces instead of forking ps per caller. `readStructuredTuiProcessIdentity` then spent its whole 5s wait inside one capture and concluded "no exact child" after a single look taken before the child existed (observed landing at ~3.5s). Absence needs a look that did not race the spawn, so require two captures before the deadline can end the loop. Both surfaced by the real-binary Claude TUI resume test, which failed ~1 in 5 under load; 14/14 now, 8 of those runs containing a capture the old 3s budget would have killed. * Let the desktop renderer negotiate Claude structured tabs The paired-client gate hides agent-session rows an agent the client cannot render. The desktop renderer's own IPC dispatches as clientKind 'runtime' advertising only agent-session.structured.v1, so the gate hid Claude rows from the surface this feature ships on. It renders them; it should say so. * Stop a slow process table from silently blinding every freshness gate Stamping `capturedAgeMs` at ps START made the number honest, and honest broke both consumers that read it. `ps -axo ...command=` measured 2.5-9.0s on an idle 2,002-process laptop and 4.0-18.6s at load 46, so the age it now reports lands past every budget: `planRelayPtySweep` refuses the stop as "too old", and the renderer's `admitRemoteForegroundEvidence` refuses the record outright. That second one is the expensive half and was outside the diff -- a refusal bumps `consecutiveInspectionErrors`, the poll scheduler backs off to its 10s floor, and agent-completion detection stops for the pane. The subsystem went blind on exactly the loaded hosts the honest stamp was meant to serve. The evidence-publishing read now gives up at 1,200ms instead of waiting out `PS_TIMEOUT_MS`. It is one budget for one question: these consumers ask whether an observation describes NOW, and past this it does not -- a late answer is refused by the age gate anyway, having first blocked a polled path for the whole capture, so a prompt `unverifiable` is both the truthful verdict and the cheap one. Both relay call sites already produce it from a rejection, and an admitted `unverifiable` costs a poll where a refusal costs the cadence. Identity proof keeps the full 15s through `getFreshProcessTableSnapshot`, because it asks whether a process EXISTS and must never read slow as absent. The budget bounds the wait, never the capture: the reader coalesces, so an abandoned wait leaves its capture running to fill the cache rather than forking a second whole-machine `ps` on the host that can least afford one. 1,200ms is bracketed rather than picked. The floor is the capture's own cost -- `command=` measured 1.15s for 1,948 processes on an idle host, and a budget under that answers `unverifiable` about a machine nobody is straining. The ceiling is the consumer's: 2,000ms, less the 500ms a TTL-shared capture may already have aged, leaves 1,500ms, and transit takes the rest. That ceiling only fits once the capture stops being charged twice. `ps` runs inside the RPC round trip, so its duration is already in `receiveDelay`, and `capturedAgeMs` is that same duration on the host's clock; summing them halved the budget this gate grants a host from ~2.0s of `ps` to ~1.0s, which is why a 1.2s capture arriving at 1.3s read as 2.5s old and was refused. Admission now takes the larger of the two. The sweep's gate keeps its sum, which is correct there: `evidenceAgeSinceListingMs` is stamped after the listing ARRIVES, so it measures planning time and overlaps nothing. A stated limit rather than an assumed one: 15s is not proven sufficient for identity proof. The same capture reached 18.6s at load 46, so that path can still time out and answer "no exact child" about a host it simply could not read in time. Narrowing it needs a cheaper question than a whole-machine argv read, not a larger number. The one test guarding this field could not fail. `beginPtyHandlerTest` installs fake timers, so `Date.now()` is frozen, the real reader reports exactly +0, and `0 <= 500` held identically for a hardcoded zero, for completion-stamping and for start-stamping -- while the real reader on that host returns thousands of ms. It now drives a measured age in and asserts the handler publishes it rather than restamping; that the reader MEASURES it correctly stays pinned separately, against a controllable clock. Both consumers get boundary coverage either side, and each new gate was ablated red before it went green. * Keep the compatibility fields off the capture the budget just abandoned inspectProcess falls back to processHasChildren and listProcesses to getForegroundProcessName, and both read the same TTL-shared capture with no budget of their own. On a slow host they joined the in-flight capture the budgeted evidence read had just given up on, so the call still blocked for the full 6-18s and the budget bought nothing -- once for inspectProcess and once per managed PTY for listProcesses. Use the degraded answers those helpers already give for an unreadable table, reached promptly. pty.hasChildProcesses keeps its unbudgeted fresh probe: it is a one-shot destructive gate that can afford to wait. --------- Co-authored-by: Merge Sim <merge-sim@local> Co-authored-by: Merge Sim <sim@local> |
||
|
Brennan Benson
|
fd9125ea8c |
feat(native-chat): Codex structured native chat restructure (#16729)
* feat(native-chat): port structured Codex sessions from restructure-recovery Rebuilds the desktop structured native-chat implementation from brennanb2025/native-chat-restructure-recovery (tip 4e31c08db3) on top of current main as a single commit, scoped to the local Codex path. Ported: - Structured agent-session core: durable record store + single-writer lease, canonical journal, agent-session wire host/attach/eviction/subscribers, `agentSession.*` RPC surface (registered via ALL_RPC_METHODS; host-side mobile allowlist included for wire compat), pty write gate, transcript additions, and the Codex app-server adapter/launch resolution. - Renderer: NativeChatStructuredSession view/composer stack, structured launch path with the single-flight guard, local structured session tabs sync, activation gate + structured inventory (read-only `agentSession.handoffStatus` probe), agent-session tabs in the tab strip, AI-vault structured session activation, and the settings pane with the parent Experimental Chat UI toggle plus the nested "Use updated structured native chat" toggle. New sessions require both flags, agent codex, no prompt, and a local non-WSL, non-Windows-host execution host (structured-native-chat-availability). - Fixes 72c013cea6 (verified Codex launch recovery), 8ddbaf5e3d (defer native terminal view switching affordances), and 4e31c08db3 (release the launch gate after a visibility retry) with their regression tests, including the third-launch-after-retry guard case. - Cross-version agent-session wire test + CI lane, packaging entries (proper-lockfile, agent-tooling asar excludes), and the wire-compat doc section. Deliberately not ported: mobile/ changes, the Claude structured runtime (only the claude-transcript-branch-proof and claude-structured-owner-identity leaf modules remain, backing the kept TUI-recovery arms), the terminal↔chat adoption/handoff flow (`agentSession.adoptTerminal`/`requestHandoff`, the handoff request engine, TUI adoption machinery, orca-runtime adoption methods), renderer switching affordances and their dead leftovers, the hook/subagent-status refactor cluster, and unrelated branch changes. The crash-during-acquisition recovery path (restart handoff adjudication, restore/reverse re-acquire, lease schema handoff keys) is kept because every plain direct launch depends on it; a trimmed handoff coordinator exposes only status/restore/close. Branch edits that targeted files main has since split (ipc/pty.ts, worktrees.ts, rpc/methods/terminal.ts, useIpcEvents, pty-connection, store/slices/terminals.ts, runtime-types, web preload) were re-applied to the split modules, preserving main's newer logic (Windows CIM fallback, browser tab close rework, cold-restore resume flow, dispatcher threading). Known seam: the mobile clipboard image-provenance CONSUMER gate ships (agentSession.send refuses unproven mobile image refs with agent_session_image_untrusted) but the producer hunk in rpc/methods/clipboard.ts stays with the unported mobile cluster, so mobile image sends into structured chat fail closed until that side ports. * fix(native-chat): trust only authenticated local image uploads * fix(build): preserve Windows process-tree patch application * test(windows): include process creation time in addon fixture * fix(build): run windows-process-tree node-gyp from the physical package dir gyp expands the node-addon-api dependency by probing node, whose cwd resolves to the package's physical directory in the store, so the emitted target is a store-relative ../../../../node-addon-api@... hop. gyp then resolves that hop against the rebuild cwd; from the node_modules symlink/junction it escapes the store and configure fails with "node_addon_api.gyp not found" (run 32999886072). Rebuild from realpath(package dir) so both bases agree, matching how the package manager itself runs native install scripts. The regression test replays gyp's expansion+resolution against the planned cwd and fails without the fix. * fix(native-chat): keep chat tabs visible through terminal closes and empty-worktree launches Two proven blockers in the native Codex tab contract: closeTerminalTab pre-empted the canonical unified close. With one terminal left it deactivated the worktree on a terminal/editor/browser-only check, blanking a workspace that still held a renderable agent-session tab; with two or more it pre-picked a successor from terminal entities only, re-stamping the group active before closeUnifiedTab's MRU/neighbor repair could land on the chat tab. Successor choice now defers to the unified contract whenever the terminal has a unified row, and deactivation is gated on the unified renderable count (matching leaveWorktreeIfEmpty), with the legacy pre-pick kept only for terminals without a unified row. A structured session created on an empty worktree was published into the host's headless group while preserveLocalLayout froze the local layout, leaving the tab in store but permanently off screen. A preserveLocalLayout owner now always takes client-owned placement — repairing a rendered leaf whose group record is missing, or materializing a rendered group on a truly empty worktree — and applies the client-derived layout repair while still rejecting host-authored layout. Regression tests drive the real store through closeTerminalTab (git worktree and folder workspace) and the real snapshot applier for the empty-worktree adoption states; all fail without the fixes. * fix(native-chat): close stale turns and retry rejected sends * fix(native-chat): retire hosted rows on structured tab activation * fix(native-chat): preserve rpc defaults across main merge * chore: format remote wire compatibility guide * test(native-chat): cover retry after unconfirmed send * fix(native-chat): reload outbox on session switch * docs(settings): disclose structured chat platform limits * fix(native-chat): await Codex launch-home preparation * fix(codex): align child-process allowlist with async trust bridge * test(identity): update inventory for tab surface refactor * fix(windows): preserve process-tree CRLF patch sources * fix(native-chat): anchor an unmatched chat echo where it was sent (#16117) * fix(native-chat): anchor an unmatched chat echo where it was sent The reported symptom was old user messages replaying below every new turn, so the conversation read as scrambled. The cause was not that the echo failed to match a transcript row. Claude consumes a mid-turn send through a `queued_command` attachment and writes no `type:"user"` record for it, so some echoes can never match, and no amount of matching will change that. The cause was WHERE an unmatched echo rendered: buildMobileNativeChatTransientData appended every pending item after the entire transcript, so it re-read below each turn that landed afterwards. Render each echo directly after the transcript row it was sent against, using the baseline the send already captures. An unmatched echo is then at worst a duplicate in the right position rather than a scrambled one, and it stays visible. Echoes sharing an anchor keep send order; a send with no baseline, or one whose anchor folding dropped, still falls back to the tail. Deliberately NOT fixed by deleting the echo. Inferring from send ordering that an echo can never match, then removing it, loses the user's own text for a message the agent did receive, and it cannot fire in the common case anyway - measured drain groups are 1,017 of size 1 against 55 larger. It also escalates an existing gap: the count pass has no baseline-tail guard, unlike the glue pass, while `messages` is a 40-row window that head-trims, resets on reconnect and grows at the front on loadEarlier, so a false landing there would license deleting a DIFFERENT outstanding message. That count-pass gap is real and left for a separate change; anchoring makes its worst case a duplicate in place rather than a scrambled conversation. * fix(native-chat): preserve folded echo anchors * fix(native-chat): preserve forward-folded echo anchors * fix(native-chat): keep leading folded echoes in place * fix(workspace-cleanup): show git status for every row (#16690) * fix(native-chat): refuse structured chat on every Windows execution path canUseStructuredNativeChat only refused win32 when a project runtime resolved, so folder-workspace keys (and other keys with no project runtime) failed open into structured chat on Windows. Fail closed on win32 unconditionally after the host check, matching the settings copy: local macOS/Linux only; Windows/WSL/SSH stay on terminal chat. * fix(native-chat): restore runtime refusals behind the win32 gate |