Reproduces stale compaction references through the current transcript-store path. Native resume of the synthetic source, fork, and repaired control succeeds, so this is not evidence of the reported context overflow.
Co-authored-by: kesiyuan <kesiyuan@bytedance.com>
* Let agents steer a running turn from send_agent_prompt and paseo send
When an agent used the send_agent_prompt MCP tool on a target that was
mid-turn, sendPromptToAgent ran without activeTurnBehavior, so the
prompt path fell through to replaceAgentRun. That cancelled the target's
active turn and every provider subagent it had in flight. The web UI,
child finish notifications and schedules already pass "steer"; the MCP
tool and the CLI had no way to ask for it.
Add an optional activeTurnBehavior input to send_agent_prompt. It
defaults to "steer" for agent-scoped callers and to "interrupt" for
top-level callers, so existing top-level behaviour is unchanged. Add
--steer to paseo send. Document both and cover them with tests.
* fix(agents): retain finish notifications during steering acknowledgement
Observe completion before dispatch and hold delivery until acceptance. A replacement discards the displaced turn observation and starts a fresh listener before its iterator runs.
* fix(agents): preserve notification ownership when follow-ups fail
Keep the previous observation until dispatch succeeds and suppress duplicate delivery during acknowledgement. Decide whether to retain the running turn from the dispatch result, including side commands requested with interrupt.
* fix(agents): deduplicate notifications across overlapping follow-ups
---------
Co-authored-by: Mohamed Boudra <boudra.moha@gmail.com>
resolveClaudeCodeVersion() probes <configured command> --version and parsed any output with the generic fallback in parseClaudeCodeVersion. With a wrapper command (e.g. cux), that output is the wrapper's own version, which then failed every minimumClaudeCodeVersion gate and silently hid Opus 5.5, Opus 5, Fable 5, Sonnet 5.5 and Haiku 5.5. Only trust output that identifies itself as Claude Code (x.y.z (Claude Code)); anything else falls into the existing unknown-version path that does not gate models.
* fix(server): return actionable OpenCode bridge errors
* fix(server): surface OpenCode tool input errors
* fix(server): surface OpenCode v2 tool errors
* test(server): cover OpenCode v2 bridge errors
* fix(server): classify missing schedules as tool input
* fix(server): surface schedule state conflicts
* fix(server): validate schedule tool inputs
* fix(server): keep tool contracts while reporting OpenCode input errors
Preserve the removal of create_agent.background and keep legacy schedule fields hidden. Verify client errors through the real bridge, catalog, and schedule service rather than mocked schedule failures.
Co-authored-by: Felipe Fidelix <felipe@fdx.is>
* fix(server): distinguish caller errors from stored data failures
Tag validation failures where tool arguments are parsed, so schema errors from persisted state remain internal failures. Preserve the missing-agent error at its producer and translate it for prompt delivery. Real HTTP regressions fail for both misclassifications before this change.
* fix(server): preserve messages for unclassified tool failures
Keep existing error messages on HTTP 500 responses so agents can correct failed tool calls. Exercise missing workspaces through the real script service and bridge route.
* fix(server): preserve clearing schedule names with blank strings
---------
Co-authored-by: Felipe Fidelix <felipe@fdx.is>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
A pending thinking change retired the live query when mode, model, or command controls requested it. Reuse the query owned by the current turn and apply pending launch settings after the turn finishes.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): run slash commands sent by another agent out of band again
* fix(omp): send /steer and /follow-up as prompts so an idle agent starts a run
* test(omp): verify idle follow-ups and parent notifications
Cover real daemon and MCP delivery, and keep the fake runtime tests independent of an installed OMP executable.
Co-authored-by: Hung Le <22406703+nhthung@users.noreply.github.com>
* fix(omp): report idle follow-up admission failures
Use the existing prompt lifecycle for an idle command so acknowledgements and late rejections settle the child and notify its caller. Keep busy commands on the streaming prompt path, and accept one runtime prompt input object.
---------
Co-authored-by: Hung Le <22406703+nhthung@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(claude): show a background subagent's action log on its card
Claude Code answers a background Agent call at launch and streams the
child's frames afterwards. Since #6067 the sidechain tracker stopped
emitting the parent card once the call had its result, so the card's
log stayed empty. The tracker now rebuilds the card from the settled
call, keeping its status, output, and label, and adds the child's
action log.
Closes#6157
* fix(claude): release a settled subagent call when the child finishes
* fix(claude): keep a background subagent's card in the turn that settled it
A background child often keeps streaming after the parent's turn ends. Its
frames then rebuilt the settled card in the next turn with an empty log, and
the projection showed it as a second card for the same call. The settled call
now rebuilds the card only within the turn that answered it.
* test(claude): model the turn boundary in background card coverage
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Generated image Markdown was concatenated into the preceding paragraph and rendered at inline size. Preserve paragraph boundaries around image output while retaining intentional inline image behavior.
Cursor agents never showed plan usage cards under their context details.
ACP sessions had no usageSession(), and the Cursor usage source only ran
global discovery, so per-agent discovery always came back empty.
Cursor ACP sessions now report a usage session (provider "cursor" plus
their launch env). The Cursor usage source resolves the cursor-agent
auth.json for that env. Sessions on CURSOR_API_KEY or CURSOR_AUTH_TOKEN,
the in-memory store, or the macOS Keychain (the default there) report
no login.
* fix(server): open worktree.terminals when a workspace is created on its own
runWorktreeSetupInBackground only opened the paseo.json worktree.terminals
when its caller passed runAutoTerminals, and only the fork-PR "allow setup"
handler did. workspace.create, the request the New Workspace screen sends,
ran worktree.setup and then skipped the terminals.
Both callers want the terminals, so the flag goes away: terminals follow a
successful setup on a newly created worktree. Reused worktrees stay excluded
by shouldBootstrap, and the agent setup continuation never reaches this
function, so nothing opens twice.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* test(server): run automatic terminal regressions in CI
Include workspace creation in the existing integration command, and append to the terminal marker so repeated command execution fails the regression.
* test(server): make workspace fixtures portable on Windows
Run the setup fixture through a command accepted by PowerShell, await its readiness, and fix the fixture Git line-ending policy. Allow Windows directory locks to settle during cleanup.
* test(server): expose setup failures and await fixture cleanup
Check the public setup status before terminal startup so failures distinguish setup errors from terminal creation. Use asynchronous directory removal so Windows cleanup retries leave the event loop available.
* test(server): start the Windows setup service asynchronously
The Windows regression starts its service but setup stays running, so terminal bootstrap is never reached. Launch the fixture service explicitly with Start-Process and retain the readiness and setup-completion assertions.
---------
Co-authored-by: David Yang <9925+davidyang@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Each streamed text or reasoning chunk was cut from the accumulated
message with slice, which can share the source string's storage. The
timeline store keeps every chunk, so a 128 KB reasoning part retained
about 128 MB: every chunk held the full message as it stood when the
chunk arrived. Long OpenCode v2 reasoning streams exhausted the daemon
worker's heap.
The timeline now emits an independent copy of each suffix, so retained
memory follows the length of the text.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Preserve the bundled Hoplite ACP adapter and provisioning transformer, activate it only after Add, and provide a searchable provider grid. Restore the contributor's original icon without the copyright mark.
Co-authored-by: Odin Glynn-Martin <odin.glynn@icloud.com>
Protocol contracts are available through the SDK peer dependency during development, but clean Git installs lack that dependency tree. Accept erased protocol type imports while retaining runtime dependency resolution and SDK ownership checks.
* fix: reject a question permission that would deliver no answer
Answering an `AskUserQuestion` permission in any shape but
`updatedInput.answers` keyed by one of its questions resolves as `allow` and
delivers nothing. The waiting agent is told "The user did not answer the
questions." — an affirmative falsehood rather than silence, so neither side can
see that anything failed. The responder sees success; the agent believes the
human declined and acts on that.
The field it is most natural to reach for cannot work. `selectedActionId` is
`z.string().optional()` with no membership check, and is read only for `plan`
kinds; a question advertises `actions: undefined`, so there is no action id to
name in the first place.
It is also unrecoverable. `respondToPermission` deletes the request from
`pendingPermissions` on entry, so the malformed answer consumes it and the retry
fails with "No pending permission request". The guard runs before that delete,
and `AgentManager` drops its own copy only once the call resolves, so both maps
survive the throw and the request stays answerable.
The answer contract is per provider, so the guard is in the Claude provider
rather than in `AgentManager`. Claude keeps an answer only when its key is a
question's full text or its header and its value is a non-empty string; Codex's
`mapCodexQuestionResponseByHeader` reads headers only and deliberately selects
each question's first option when nothing maps, which a shared guard would have
rejected; OpenCode reads headers only as well.
`resolveClaudeAskUserQuestionAnswers` is extracted from
`normalizeClaudeAskUserQuestionUpdatedInput` with no change in behaviour, and
both it and the guard now use it. So "would this deliver anything" is answered
by the code that does the delivering, rather than by a second copy of its rule —
which matters more than it looks, because that rule has two details easy to
restate wrongly: `readNonEmptyString` tests `value.trim()` for emptiness but
returns the *original* string, so a question named `" Which colour? "` is keyed
by that padding; and an empty `updatedInput.questions` array is a list that wins
over the request's own questions and maps nothing, not an absent one.
Verified on a live daemon over the MCP `respond_to_permission` path against a
seat whose question text is literally `" Which colour? "`: `selectedActionId`,
the trimmed key, an empty echoed question list and a numeric answer are each
rejected with the request still pending, and
`{"answers":{" Which colour? ":"Ochre"}}` then reaches the agent.
One known limit: the WebSocket path (`session.ts handleAgentPermissionResponse`)
catches the throw and emits an `activity_log` error rather than failing the
call, because a permission response is modelled as a notification. So
`paseo permit allow` still prints `allowed` and exits 0 for a rejected shape.
The destructive half is closed on every path — the request survives and the
agent is never told a falsehood — but closing the reporting half needs a
protocol change.
* test(claude): cover the question-answer-required retry path
A non-deliverable answer to an AskUserQuestion permission is rejected, but
the pending request survives it: a corrected retry against the same request
id is accepted, and the waiting agent receives the answer.
Mutation testing on a fork found that the existing suite does not pin this
ordering — moving the deliverability check to run after the request is
deleted from pendingPermissions still leaves every existing test green. This
test reds against that change.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---------
Co-authored-by: Thomas Van <minhthong@gmail.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: paseo-bot <paseo-bot@users.noreply.github.com>
* fix(server): name an imported Claude agent after its /rename
Importing a Claude session named the agent after its first prompt even when the
session had been renamed in the CLI, so the name had to be set again by hand.
#4216 fixed the Import Session listing; the imported agent still kept the
prompt.
Read the last usable `custom-title` from the transcript and pass it as the
imported config title. The record is re-emitted on every save, so searching
backwards from the end of the file finds the current name without walking a
transcript that can reach tens of megabytes.
Only an explicit rename is carried over. Agents are named after their first
prompt line since #1563, so adopting the generated `ai-title` here would put
that LLM summary back.
Refs #2513
* refactor(server): drop unused generality from the Claude title reader
readLastClaudeTitle took content, type and field so it could read both
custom-title and ai-title records, which the listing needed before #4216 took
that over. Only the custom-title read is left, so the two extra parameters are
dead flexibility that also broke the single-object-parameter rule in
packages/server/CLAUDE.md.
Collapse it to readLastClaudeCustomTitle(content) and pull the record type into
a constant so the needle and the type check share one literal.
Hooks skipped: the repo-wide typecheck fails on two pre-existing Electron
typing errors in packages/desktop/src/main.ts that reproduce with this change
stashed. packages/server typecheck, lint and format all pass.
* test(server): drive the real resume path in the import title tests
The tests assigned an inline arrow function to client.resumeSession to keep
Claude Code from spawning. Importing replays the transcript from disk and never
starts a process, so the stub was never needed. Dropping it runs the real
resume path, which is what docs/testing.md asks for, and removes the arrow
assignment that docs/coding-standards.md rules out.
Hooks skipped: the repo-wide typecheck fails on two pre-existing Electron
typing errors in packages/desktop/src/main.ts that reproduce with this change
stashed. packages/server typecheck, lint, format and the 83 claude tests pass.
* fix(server): narrow the transcript read error instead of asserting it
The ENOENT check cast the caught value to NodeJS.ErrnoException to reach
`code`, which is the kind of assertion docs/coding-standards.md rules out:
narrow with `if`, do not silence the compiler. Add the same local type guard
pid-lock.ts already uses and test through it.
Behaviour is unchanged. A value that is not an Error still falls through to
the rethrow, exactly as reading `code` off the assertion did.
Hooks skipped: the repo-wide typecheck fails on two pre-existing Electron
typing errors in packages/desktop/src/main.ts that reproduce with this change
stashed. packages/server typecheck, lint, format and the 83 claude tests pass.
* fix(server): let claudeProjectDirSync resolve the Claude config dir
The import title reader built the config dir itself through os.homedir(),
but os is no longer imported here, so the fallback threw a ReferenceError
whenever CLAUDE_CONFIG_DIR was unset. claudeProjectDirSync already resolves
the same default.
* test(server): give the import title helpers object parameters
* fix(server): preserve imported session renames across history locations
The import title reader must use the same provider environment and transcript
lookup as history replay. Otherwise an imported session has its history but
loses its explicit name when its configuration directory or project changed.
Co-authored-by: platanus-kr <platanus.kr@gmail.com>
---------
Co-authored-by: platanus-kr <platanus.kr@gmail.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
ACP command updates discarded input hints before the autocomplete command list reached consumers. Preserve the advertised hint, retaining the empty hint for commands without input metadata.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: bwestlund17 <brian.westlund@proton.me>
* fix(gitea): prefix fork PR branches with the owner
Two forks with the same branch name (say patch-1) ended up on the same
local branch. The gitea service had no buildPrLocalBranchName or
defaultCheckoutRefs, github has both.
getPullRequestCheckoutTarget also fell back to refs/heads/<name> on
origin for fork PRs. Thats the base repo's branch, not the fork's, so
a failed refs/pull fetch checked out the wrong branch without an error.
fixes#3515
* refactor(forge): share fork branch naming with github
moved the owner normalizer and the branch name builder to
utils/change-request-checkout.ts, gitea and github both use it.
github changes a bit: a fork PR with an unknown or unusable owner (eg
deleted fork) now gets pr-N/<branch> instead of the bare branch name,
which could collide with a local branch the same way #3515 did.
* fix(gitea): PRs from a deleted fork are still cross repo
the API drops head.repo once the fork is gone, we read that as a same
repo PR and fetched refs/heads/<name> from origin
* fix(gitea): handle AGit PRs and deleted head branches
neither has a branch on origin. forgejo reports refs/pull/N/head as
head.ref for both, so skip the refs/heads fallback and dont track
origin.
also try refs/pull/N/head on upstream for clones where origin is your
fork, like github does. that list lives in the shared util now.
* test: resolver harness uses the real branch helper
plus deleted forks and same repo PRs without an origin branch
* fix(gitea): collapse pull-ref-only heads to pr-N in local branch names
buildForkLocalBranchName still used the raw headRef even when it was
just the pull ref (refs/pull/N/head), which agit prs and deleted head
branches report as their headRefName. that gave local branches literal
names like refs/heads/refs/pull/3/head.
* fix(gitea): surface hasHeadBranch on the pr checkout target
resolveForkCheckoutFacts already knew whether a pr had a real head
branch (agit flow, or head.ref == the pull ref both mean no). now that
signal reaches buildPrLocalBranchName so worktree names and local
branches stop being named after the raw pull ref.
covers same-repo, fork with owner known, and deleted fork all with no
real head branch. added an end to end gitea-service test plus resolver
cases since the worktree slug follows localBranchName too.
* refactor(forge): one head ref kind on the checkout target
hasOriginHeadBranch and hasHeadBranch could describe contradictory
states (cross-repo checkout claiming an origin head branch). collapse
both into headRefKind: "branch" | "pull-ref" and derive origin
tracking in the resolver instead of storing it separately.
Co-authored-by: Alexander Adam <alexander.adam@vade.io>
* fix(worktrees): avoid branch namespace conflicts in fork checkouts
Fork owner prefixes can name an existing branch as a parent, which Git rejects even when the full proposed name is unused. Allocate names against parent and child refs, suffixing the conflicting segment while preserving existing branches and fetched commits.
* test(worktrees): make fork fixture line endings deterministic
Windows Git converts the fixture checkout to CRLF when core.autocrlf is enabled. Set the repository fixture to preserve its written LF bytes so the content assertion checks the fetched patch on every platform. Reproduced the failure locally with inherited core.autocrlf=true and verified the corrected setup against that same configuration.
---------
Co-authored-by: Alexander Adam <alexander.adam@vade.io>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(claude): keep context meter on the active model
Claude result modelUsage accumulates models across the session. Reading its largest window replaced Haiku capacity with a previous Opus window at turn completion. Select the runtime model entry and retain the seeded limit when that entry is absent or invalid.
* refactor(claude): name context fixture inputs and simplify result usage
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* test(server): stop expecting request-derived upload IDs
* test(server): check full uuid shape for upload id
* test(server): format generated upload ID assertion
Retain the contributor correction to the stale request-derived ID expectation while satisfying the repository formatter.
Co-authored-by: hadan8977 <zc-huang23@mails.tsinghua.edu.cn>
---------
Co-authored-by: hadan8977 <zc-huang23@mails.tsinghua.edu.cn>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(git): preserve renames in committed file diffs
Follow the requested destination through rename detection so the patch agrees with commit history instead of showing the whole file as newly added.
* test(git): keep rename fixtures valid across platforms
* fix(git): keep added copies out of committed rename diffs
Detect rename sources across the first-parent commit before narrowing the patch. Follow mode also detected copies, causing added files to open as partial diffs against unchanged files.
* perf(git): share committed rename discovery across file requests
Opening a commit requests all of its file diffs concurrently. Share pending and completed rename scans for immutable commit IDs in a bounded cache, while symbolic refs always read fresh metadata.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(git): return checkout diff hunks from subdirectory workspaces
Checkout paths are repository-relative, but one-shot diff reads used the workspace subdirectory for Git pathspecs. Resolve the worktree root once so patch selection and file reads use the same coordinates as the live subscription.
* test(server): await bounded discard fixture cleanup on Windows
The discard assertions pass but Windows can retain a directory lock during cleanup. Await asynchronous removal with bounded retries after daemon shutdown, preserving failure for persistent locks. Matches the verified cleanup correction in #6912 for #6918.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): report a terminal whose command cannot start instead of leaving a dead tab
Opening a terminal profile whose command is not installed left a tab that could
never be closed. Two problems sat on the same path.
A terminal that dies during creation was resurrected by its own create. The
worker sends terminalCreated, the createTerminal response and terminalExit in
the same tick, so the parent drains all three before the awaiting create
continuation resumes; that continuation called registerRecord a second time and
put back a terminal the exit had just removed. Removal is owned solely by the
terminalExit event and that event never comes twice, so the record outlived
every close attempt: getTerminals kept listing it, capture returned nothing,
kill was a silent no-op that still reported success, and only a daemon restart
cleared it.
The parent now tracks the ids with a create in flight and keeps the finished
record when an exit removes one of them, so the create continuation returns the
exited session instead of registering a live one. Removal still happens exactly
once, in the terminalExit path.
With the stuck tab gone, a missing command produced nothing at all: no tab and
no message. node-pty spawns directly through execve on POSIX, with no shell, so
a command that does not exist fails inside the spawn helper and surfaces as exit
code 1 with no output -- exactly what `false` produces, so the exit carries no
usable signal. createTerminal now resolves the command before spawning and fails
the create when it cannot be found, so the daemon answers create_terminal with a
null terminal and "<command>: command not found", which the app already renders.
A command that starts and then exits non-zero still opens its tab.
That resolution has to match what execve will do in the child, or a working
profile is refused: it uses the terminal's own environment, built first so the
check sees the per-cwd registered PATH and the entries buildTerminalEnvironment
prepends, and it resolves a command containing a separator against the
terminal's cwd rather than through PATH. Windows keeps resolving inside
resolveTerminalSpawnCommand, which has its own shim handling and fallback.
* fix(terminal): check exact executable names and verify launch output
---------
Co-authored-by: KiraPC <pasqualecarmine.carbone@gmail.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(git): discard repository paths from the worktree root
* test(git): cover discard from nested workspaces through daemon RPC
* test(git): close the daemon before removing discard fixtures
Windows CI reaches the discard assertions but cannot remove the observed
repository while the test daemon is alive. Run fixture removal after the
daemon's afterEach cleanup, preserving all behavioral assertions.
Co-authored-by: 见安 <73627805+Mar-garet@users.noreply.github.com>
---------
Co-authored-by: 见安 <73627805+Mar-garet@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* Let readFile callers disable the request timeout
readFile resolves only when FileEnd arrives, so the default 60 s request
timeout bounds the whole transfer. A large file over a slow relay fails
before it finishes even though every chunk is arriving. Thread an optional
timeout through requestFileExplorer; 0 disables the timer, as the
workspace.create request already does, and a dead connection still rejects
on disconnect.
The relay e2e case reads a 1 MiB file through a real daemon and a local
relay with timeout 0 and compares SHA-256 digests with the source.
* Reject a binary file transfer whose byte count differs from FileBegin.size
concatByteChunks allocated FileBegin.size bytes and copied the chunks in:
a short transfer resolved as a zero-padded file, and a long one threw a
RangeError inside the transport callback and left the request pending.
Count the received bytes at FileEnd and reject the read when they differ
from the announced size.
* Download files over the session socket when no direct connection is active
The download store only knew the HTTP token route and picked the first
directTcp connection in the host profile, whatever connection the host
runtime was using. A relay-only host has none, so every download ended in
"Download host is unavailable." (#543). SSH, socket and pipe hosts failed
the same way, and a stale directTcp entry won over a working relay (#3753).
Choose the transport from the active connection: directTcp keeps the HTTP
token download unchanged; anything else reads the file over the session
socket with the binary frames previews already use, then hands the bytes
to the browser's download manager on web and Electron or writes them to
the cache directory and offers the share sheet on native. Files over
64 MiB are refused before any byte moves, using the existing fileTooLarge
string, and a daemon response without content fails instead of saving an
empty file.
* fix(app): enforce the session download size limit during the read
The token size is sampled before the file is read. Pass the download byte budget through the actual read so a changed file cannot bypass the in-memory limit.
Co-authored-by: Michael Vandesompel <mv@droggol.be>
Co-authored-by: YASICJUNWOO <joonoo3@naver.com>
Co-authored-by: JrDw0 <JrDw0@users.noreply.github.com>
* test: verify TCP downloads and own relay fixture cleanup
---------
Co-authored-by: Michael Vandesompel <mv@droggol.be>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: YASICJUNWOO <joonoo3@naver.com>
Co-authored-by: JrDw0 <JrDw0@users.noreply.github.com>
* fix(server): start newly configured workspace services without restarting
Extend the workspace port plan when paseo.json declares another service, preserving existing peer ports. Serialize extensions and roll back only their new reservations on failure.
* fix(server): serialize service restarts with port plan extensions
Keep concurrent restart and newly added service allocations in the same workspace update transaction so neither overwrites the other.
* fix(server): keep service port update failures scoped and release cancelled reservations
* test(hub): show creation errors in re-enrollment failures
Assert the error field along with success so failed CI reports include the reason agent creation was rejected.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(client): reject a message over the daemon's size limit instead of reconnecting in a loop
The daemon closes the WebSocket with 1009 when a message exceeds its 100 MiB
limit. A workspace or agent creation that loses its connection mid-request
stays pending and, after reconnecting, finds no record of itself on the
daemon (the oversized frame was never read), so it resubmits the same frame
and the connection drops again every couple of seconds until the app reloads.
The client now checks each outbound frame against the same limit, which
protocol exports and the daemon sets explicitly, and fails that request with
a size error while the connection stays up.
Refs #5501
* test(client): cover the UTF-8 byte count at the message size limit
* fix(client): reject oversized relay messages without disconnecting
Check the encrypted wire size synchronously before submission, so a frame below the direct limit cannot close the relay and leave workspace creation resubmitting forever. Keep the relay limit shared with the existing checkout diff budget.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* Show Claude option previews in the question card
* Wrap long lines in question option previews
* Check option previews under the option that owns them
* Show the whole option preview on native so every line is reachable
On Android, a slow drag inside a preview capped at 240px did not scroll
it, so lines past the cap could not be read. The preview's nested
scroll sits inside the option's Pressable; the same drag scrolls it
once the preview is outside the Pressable. Native now renders the
preview at full height so the chat scroll reaches every line, and a tap
on it still selects the option. Web keeps the 240px cap.
The mock agent's rollout preview is now taller than the cap so the
Android replay can check the last line.
---------
Co-authored-by: Gabriel Giordano <hello@gabrielgiordano.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(omp): show manual compaction progress and refresh context usage
OMP answers a manual compact with its RPC response alone, without
compaction_start or compaction_end, so /compact on an OMP agent showed no
row and left the context meter at the pre-compaction value until the next
turn. The compact request now emits the loading and completed rows itself
and publishes the session's context usage once it returns.
The fake OMP runtime no longer emits compaction events for a manual
compact, matching what OMP sends.
Closes#5966
* fix(omp): keep a running turn's usage polling through a manual compact
Refresh the compacted context usage without ending the active turn's
poll, wait for the usage update in the real OMP test, and cover a failed
manual compact.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): accept voice and model names from self-hosted speech servers
features.voiceMode.tts.voice and TTS_VOICE were validated against the
OpenAI voice catalog, so a server whose voices live under other names
could not be configured at all: config.json failed to parse and an env
var threw inside the speech config resolver, which config.ts calls
without a catch, so the daemon refused to start. TTS_MODEL had the same
problem for model ids the server serves itself. Both are free-form now,
with model ids keeping their case because rosters use values like
kokoro-v1:Q8_0. Regenerate the config JSON schema.
* fix(server): send TTS voice names as written and keep the config schema current
A self-hosted OpenAI-compatible server owns its voice ids and their case, so
lowercasing the configured voice could turn a valid id into one the server
does not know. Keep the published config schema in its existing form and
change only the voice field.
Co-authored-by: khimaros <me@khimaros.com>
* fix(server): drop the TTS baseUrl from the init log
The configured baseUrl can carry a token in its query string, and logging
it is not needed to accept self-hosted voice and model names.
* test(server): await bounded worktree fixture cleanup on Windows
---------
Co-authored-by: khimaros <me@khimaros.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): keep a symlinked config.json linked when saving
Paseo saved config.json by renaming a temporary file over its path, which
replaced a symlink with a regular file and left the linked file unchanged.
Private file writes now resolve the symlink and replace the file it points at.
Closes#5932
* test(server): check a symlinked config.json stays private
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Adds a regression test for #5928: a child that finishes its initial turn before its turn start returns still sends one finish notification to its parent. The fix itself is on main from 0009a7e.
Co-authored-by: Ahmet İzgi <ahmetizgi@gmail.com>
* fix(claude): rewind a session that lives in the provider's CLAUDE_CONFIG_DIR
Rewinding the conversation forks the Claude session through the SDK's
forkSession, which looks for the transcript under the daemon's own
CLAUDE_CONFIG_DIR or ~/.claude. A provider whose env sets CLAUDE_CONFIG_DIR
has its transcripts elsewhere, so every rewind failed with
"Session <id> not found".
Fork through a session store rooted at the folder of the transcript the
agent already resolves with the provider's config dir, so the SDK reads
the session where Claude Code wrote it and writes the fork beside it.
Closes#6321
* fix(claude): pass only transcript entries to the rewind session store
The session store parsed each transcript line and cast it to a session
entry without checking it. Keep only JSON objects with a string type, and
cover transcripts with non-entry lines in the real-SDK rewind test.
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(codex): split app-server stdout on newline only, not U+2028/U+2029
node:readline treats U+2028 and U+2029 as line terminators, but the Codex
app-server writes them raw inside JSON strings. Any message containing them
(e.g. tool output or code that mentions JS whitespace) was split into
fragments that failed JSON.parse and were dropped as "non-JSON stdout lines".
Dropped live events could leave an agent stuck in "running", and a
thread/resume response containing them never resolved, leaving the UI on
"Updating messages" indefinitely.
Replace readline with a newline-only line reader (StringDecoder for UTF-8
chunk boundaries) and add a regression test.
* fix(codex): read large app-server stdout lines in linear time
The newline-only stdout reader rescanned the whole accumulated line on
every chunk, so a multi-megabyte thread/resume response took seconds to
frame. Search only the newly decoded text and join the unterminated
parts once the line ends.
Co-authored-by: James Tsang <wtzeng1@gmail.com>
* test(codex): check the large chunked line without a wall-clock limit
Co-authored-by: bebeebebbebebbe <54549544+kotusenka@users.noreply.github.com>
---------
Co-authored-by: Graham <graham@eger.dev>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: James Tsang <wtzeng1@gmail.com>
Co-authored-by: bebeebebbebebbe <54549544+kotusenka@users.noreply.github.com>
* fix(server): allow worktree creation after Git initialization
Fresh project checkout reads could publish Git capability while worktree creation retained a cached non-Git snapshot. Refresh an existing snapshot when the fresh read observes a classification change, before returning the new capability.
* test(server): cover Git initialization in the standard integration suite
* test(server): separate stress-test writes from deletion delivery
Wait for active classifications and queued deliveries between the distinct write and deletion phases. Otherwise a late create and delete coalesce as update, leaving the strict deletion assertion waiting indefinitely. Preserve every event assertion without waiting for scheduled native safety audits.
* test(server): compare canonical repository paths on Windows
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Seed active usage from the explicit 1M modifier and the model resolved by Claude Code initialization. Preserve SDK-reported limits on repeated initialization and reset them when the selected model changes.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(server): release ACP provider process after history resume
Resuming a persisted agent with purpose "history" only replays the
session into the canonical timeline, but ACP sessions kept the spawned
CLI process resident afterwards. Every client that fetches an archived
agent's timeline (the app reconnects fetching every visible agent)
therefore re-materialized a provider process per archived agent and
nothing ever released it — measured ~280 MB resident per dsh agent.
After the replay and configured overrides finish, a history-purpose
resume now settles the commands gate, drops the connection, and
terminates the child. The session object stays registered with the
replayed history so read paths keep working; the existing
`this.child !== child` exit guard already ignores the released
process, and interactive use goes through reloadAgentSession which
spawns a fresh process. This matches the runtime-only contract the
Codex provider already implements.
* fix(server): retain child refs when history-runtime release times out
Addresses Greptile review: if terminateProcess returns kill-timeout the
child stays registered so close() can retry instead of leaking an
untracked process. The test now asserts the process is already dead when
resumeSession resolves, which pins the documented release-before-return
contract instead of a post-hoc poll.
Greptile iteration 1.
* fix(highlight): render Astro diffs containing empty expressions
Empty Astro expressions produced zero-length nested JavaScript ranges, which rejected the whole diff. Leave those expressions in the markup tree and preserve highlighting around them.
* test(highlight): make empty Astro expression assertions explicit
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(client): restore provider readiness on mobile
React Native uses an AbortSignal without throwIfAborted. Check aborted with the existing API disposal error so readiness works while preserving cleanup and cancellation.
Co-authored-by: GeorgeR <897457+GeorgeR@users.noreply.github.com>
* test(plugins): run mobile readiness regression in CI
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: GeorgeR <897457+GeorgeR@users.noreply.github.com>
* fix(server): reach IPv6-only workspace services
Workspace service forwarding and health probes forced IPv4 loopback,
returning 502 and unhealthy for a reachable IPv6-only service. Select
between both literal loopback addresses for HTTP, upgrades, and probes,
without relying on the system's localhost DNS entries.
Co-authored-by: Marcelo Barbosa <celocb@gmail.com>
* test(server): keep proxy POST mechanics in a helper
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Marcelo Barbosa <celocb@gmail.com>
Publish the existing thinking-option event as soon as Pi changes its effective level, keeping subscribers and persistence in sync before the next state refresh.
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
