A session created in a git worktree while the client was already running did
not render: the message list stayed empty while the prompt and the assistant
reply were both present in the session, visible on any fresh load. Reported as
prompting in a worktree sometimes not working.
Ownership was read from which child store holds the session. That is
containment, not ownership. A project's session list includes the sessions of
its worktrees so the sidebar can group them, so the parent repository holds
worktree sessions too, and whichever store bootstrapped first won. Captured
mid-failure, the two signals disagreed outright:
owningDirectory /repo <- parent, merely holds it
recordDirectory /repo/.worktrees/feature <- the session's own directory
The parent won, so every fetch was addressed to a directory that does not own
the session, the session id resolved to undefined there, and the requests
failed as /api/session/undefined in a retry loop. The session's own record is
now believed; store membership remains the fallback for a record that carries
no directory.
This also explains why the previous commit alone was not enough: settling the
guessed directory adopted this same wrong value and then cleared the guess,
which prevented any later correction.
Verified against the reproduction rather than by reasoning. Before: three of
four runs never rendered. After, on a clean build with the instrumentation
removed: three of three rendered the reply live, each routed to its own
worktree. Tests cover ownership disagreeing with containment, plus both
directions of the guess promotion.
Selecting a session whose directory this client has not indexed yet routes it
through the active directory. That is a deliberate, documented guess: it keeps
routing usable while the owning store bootstraps, and it is excluded from both
the resolver and persistence.
Nothing settled the guess afterwards. `setSessionDirectory` performs exactly
that promotion, but only confirmed destinations call it — a completed move or a
worktree this client created. A session whose directory the client learned about
later, such as one in a worktree created outside this client, kept the guess
forever: every message fetch was addressed to the parent repository, which does
not own the session.
Captured for such a session before this change, with the session already
indexed and its owning store known:
routedDirectory .../worktree/feature
currentSessionDirectory /repo <- guess, never settled
opencodeClientDirectory /repo
conflict selected -> /repo
and after:
routedDirectory .../worktree/feature
currentSessionDirectory .../worktree/feature
opencodeClientDirectory .../worktree/feature
conflict null
Directory bootstrap completion is the moment the authoritative directory first
becomes readable, so the promotion runs there. It only ever promotes a guess:
a confirmed selection and a selection that has since moved on are both left
alone, and tests cover both directions.
This removes a real routing split-brain. It does not by itself fix the reported
symptom of a session created mid-session never rendering; that remains open.
A queued message is removed from the queue only after its send resolves,
so between dispatch and resolution it stays visible to every reader — and
a composer submit merges the whole queue into its own send. Over a relay
that window is seconds, long enough to deliver the same message twice.
The queue now tracks which entries are awaiting the server. Dispatchers
skip them, clearQueue retains them so the pending send can still remove
or restore its own entry, and the flag is not persisted because a restart
has no in-flight sends.
A prompt whose response is lost after the request left the client may
already be running server-side. The relay tunnel reported those failures
as plain text errors ("stream aborted by host", "relay keepalive
timeout"), which matched none of the patterns in isAmbiguousSendFailure,
so an accepted prompt was rolled back and the message queue re-sent it —
two independent AI responses for one user message (#2425). Direct
connections never hit the path.
Transports now tag dispatched-but-unconfirmed failures and the classifier
reads the tag before falling back to status/text heuristics. Confirmation
waits for the connection to actually return (bounded) and retries with
backoff instead of two attempts 150ms apart over the just-broken tunnel.
Expanding a project with more worktrees and sessions than MAX_DIR_STORES put
the sidebar into an endless request loop (#1472).
Every sidebar row calls ensureChild during render, but the pin that protects
the directory is only taken in an effect after commit. ensureChild marked the
directory and ran eviction synchronously, so directories that were actively
rendering looked unpinned and were disposed. The next render recreated them
with a loading status, which issued another bootstrap request, and the cycle
repeated for as long as the project stayed expanded.
Raising the limit only moves the cliff, so the limit is now a soft target
instead: a directory touched within a grace window is never an overflow
victim. A burst of live directories overflows the cache briefly rather than
thrashing, while idle-time eviction still bounds it. Eviction is also coalesced
into one deferred pass per tick, so a render that mounts many rows no longer
sorts and scans every directory once per row, and a whole commit's pin effects
settle before anything is considered for disposal. Releasing the final consumer
stays synchronous, since that is an explicit lifecycle edge.
The idle profiler gains --expand-projects to reach this state.
Not yet verified end to end: reproducing the loop needs many worktrees under
one project, which this development environment does not have.
`getRuntimeKey` keys caches, stores, and persisted state across the whole UI,
so it runs on store reads, event handling, and render paths. Until the runtime
endpoint is explicitly initialised, every call re-derived the key by trimming
two injected globals and constructing three URL objects.
In a streaming capture this made `readInjectedLocalOrigin` the single most
expensive application function: 315 ms of self time, 12% of all main-thread
busy time. After the change it does not appear in the profile at all, and the
same capture went from two long tasks to none, with the longest task dropping
from 210 ms to 47 ms.
The key depends only on the active API base URL and two injected globals, and
`switchRuntimeEndpoint` writes the injected API base URL at runtime, so the
cache is validated against the raw untrimmed values rather than memoised
outright. That comparison allocates nothing and still recomputes as soon as any
input changes. Tests cover both directions, including an operation-count
assertion that repeated calls construct no URLs.
The streaming profiler also reports output-normalised metrics, because response
length varies between runs and makes per-second totals incomparable.
Use getRequestDirectory and x-opencode-directory like the other skill
mutations, and pin renamable list/store mapping with focused tests.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Keep the active todo text hidden below 38rem, as before, while hiding only the changed-files secondary label below 30rem. This fixes the mobile collision without reintroducing the documented narrow desktop overlap.
Session directory resolution had no precedence contract: the selection-time
directory short-circuited every lookup, and a persisted runtime value was
consulted before the authoritative record. A worktree session selected before
its directory store bootstrapped kept the active-directory fallback, and that
guess was persisted, so it survived reloads and restarts.
Directory resolution now lives in one module and orders sources by whether the
server confirmed the path, not by whether the value is local or synced:
authoritative (the child store that holds the session) > server-confirmed
selection > worktree attachment/metadata (the requested path, pre-canonical) >
remembered. A guessed selection is no longer persisted, remembered, or ranked.
Chips read the same resolution the composer used, so queue keys cannot diverge.
Queued auto-send could strand an item indefinitely: backoff, missing send
configuration, and the recent-abort window all returned without scheduling a
wake-up, so the queue only retried when an unrelated status or directory change
re-ran the effect. A retry scheduler now wakes it at the earliest known time.
A rejected send rolls the optimistic message back while the composer stays
silent for transport failures, which makes it indistinguishable from nothing
happening. Failures are now recorded to a bounded in-memory log surfaced in the
About diagnostics report, alongside a directory-resolution breakdown, plus
__opencodeDebug.diagnoseSessionDirectory() and getRecentSendFailures().
Prompted by a report of worktree prompting silently failing. That failure was
not reproduced locally, so the diagnostics are what will identify it.
Terminal creation no longer waits for the Ghostty viewport to report its
size: it starts the PTY immediately with a container/font-derived
provisional size (falling back to 80x24), then resizes once the real
viewport dimensions are known, with a dedupe guard while sizing settles.
Starting the shell earlier means it can emit device/theme queries before
a browser terminal is attached to answer them, so the server now answers
primary device attribute queries itself (Fish blocks ~10s on this at
startup) and bun-pty buffers output emitted before a data subscriber
attaches. Also fixes a few WebSocket transport reconnect races surfaced
by session creation now overlapping renderer setup.
Expose authoritative renamable on skill list responses using the same
managed-root policy as renameSkill, drop the divergent UI path heuristic,
and remove an unused rejection-test fixture.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Merge of main brought the de locale without the custom-provider keys
added on this branch, which broke the UI build type cast to I18nDictionary.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Restrict in-place skill rename to managed skill directories, require
frontmatter name to match before moving, roll back/reject with tests,
hide rename in the UI for unmanaged paths, and drop unused toast keys.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
On narrow mobile widths the chat status row rendered the edited-files
summary, diff totals, todo trigger, and chevron in a single flex line.
The left side has several fixed-width parts (file count, +/- counts,
chevron) so it could not shrink past the right side and the right-side
controls visually overlapped the truncated 'changed in workspace' label.
Fold the existing 24rem hide rule for status-row__changed-label into the
same 30rem container query that already hides status-row__active-todo,
and lower the breakpoint from 38rem to 30rem so the secondary label
remains visible on a wider range of mobile widths while the two
non-collapsible sides stop competing for space.
Validation:
- bun run type-check (all workspaces, exit 0)
- bun run lint (all workspaces, exit 0)
- mobile HMR route (HTTP 200)
- visual: before/after screenshots at ~360-400px width
performConfigRefresh passed client-directory-first path into
invalidateSkillsLoadCache, missing the active-project cache key used by
loadSkills after the repository-local skills discovery fix.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Derive the effective OpenCode config layer (custom > project > user) from
provider sources and send it on PUT /api/provider so project/custom edits
update that layer instead of creating a global user override. Resolve
OPENCODE_CONFIG at call time and add UI/web/VS Code coverage for scoped
upserts.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
Rename skills by moving the skill directory and updating frontmatter
name instead of recreate-with-stub-description, which wiped the body
and supporting files.
Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>
A guided explanation is only useful in a language the reader reads, so the
panel header gets a language picker alongside the model one, defaulting to
the interface language. Like the model, it is request state rather than a
setting: the language travels with the read and the generation, and the one
a walkthrough was written in is stored with it, so reopening a review
describes what is there instead of what a fresh one would be.
Only prose is translated. Hunk aliases resolve back to hunk ids and
icon/importance are validated against fixed English values, so a translated
one would be dropped by the normalizer — silently losing an anchor or a
style. Identifiers and paths stay as they appear in the code.
The language is part of the cache key, and a read now asks the cache for the
exact request it was given before falling back to the pointer. Without that
the panel answered a request to switch languages with the text it already
had, leaving the other language unused in the cache.
Alongside it:
- The answer budget is derived from the resolved model instead of a flat 24k.
That number was the same for a 64k-context model and for one that admits to
384k output tokens, and on the latter it was the only reason generation
failed: the model spent the whole allowance reasoning and returned nothing.
It is now min(96k, max(24k, a quarter of the context)) capped by the
catalog's output limit, decided once so the input reserve and the request
cannot drift apart.
- A read no longer offers Cancel. It is a few hundred milliseconds of git with
nothing to cancel, and the button flickered on every model or language
change. When the panel is showing a fallback, a banner names what is on
screen versus what was asked for — only once the read has settled.
- The header keeps one 32px control height and drops its labels below 680px
instead of squeezing them to two letters and an ellipsis.
Docs and module documentation updated in every locale.
Use the real status-warning tokens (--status-warning-bg did not exist, so
the card rendered without a fill), drop the decorative icons, and fold the
conflict count into the title as a single full-message key per locale.
The operation description now wraps instead of truncating, and both the
conflict and ready-to-continue states share the same action layout.
Follow-up to #2574 and f95f1ab18, which guarded the archive path. The
delete path had the same two defects and worse consequences.
`useSessionUIStore.deleteSession`/`deleteSessions` declared an `options`
parameter and discarded it on both paths, so a caller-supplied runtime
key was a silent no-op. `SessionDialogs.tsx:416` already passes options
today and they never reach the action.
The delete path also never rechecked the runtime. Session IDs are not
unique across runtimes, so a response produced by a previous runtime
could commit `finalizeConfirmedSessionDeletion` against the runtime the
user switched to: evicting an unrelated session from the live and global
stores and calling `cleanupPersistedSessionState`, which erases queued
messages, todos, folder membership, inline-comment drafts, chat draft,
and pins. That is user data loss, not stale cache.
`cleanupPersistedSessionState` already rejects an identity whose runtime
is no longer active, but `finalizeConfirmedSessionDeletion` defeated that
check by passing the live `getRuntimeKey()` at commit time, comparing a
value with itself. It now forwards the captured key.
Adopt the default-on shape from f95f1ab18: `expectedRuntimeKey` defaults
to the active runtime in `deleteSession`, `deleteSessionInDirectory` and
the new canonical `deleteSessions` action, and is rechecked before the
request and before every reconciliation. A `404` still means "already
deleted" and commits cleanup, but only while the captured runtime is
active; after a switch it describes the wrong runtime and the action
reports failure instead of committing.
Also documents the throw contract of `patchSessionMetadata`, a
non-blocking nit raised by the review bot on #2574.
Bohdan Triapitsyn