custom
9
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Bohdan Triapitsyn
|
85c4320825 |
Settings storage with scopes, and project setup that can live in the repository (#3413)
* refactor(settings): settings registry and intent-gated writes
Problem: every setting lived in a flat document with ten hand-maintained
key lists that had drifted (three keys the server silently dropped, five
it kept that nothing read), and three code paths wrote to the server
without a person changing anything: the theme persist effect on mount,
bootstrap seeding of server-missing keys, and the auto-save echoing
values just adopted from the server.
Approach: one registry (packages/ui/src/lib/settings/registry.ts) names
every key with its scope (instance / profile / device), a boundary parser
and its store binding; DesktopSettings, the sanitizer, the mirror, the
apply step and the auto-save derive from it. A generated JSON snapshot
carries the key list to the server and the VS Code bridge. Writes carry
intent: the theme context writes only from its user-facing setters, a
missing server key leaves the local store alone instead of resetting it,
updateDesktopSettings drops values the server already holds, and the
auto-savers treat values applied from the server as a new baseline.
Testing: bun test packages/ui (registry + persistence suites cover zero
writes on load, dedup, toggle-back cancellation, failed-save retry, and
snapshot freshness); tsc for every workspace.
* refactor(ui): read and write settings through the shared path only
Problem: fourteen pages and stores fetched /api/config/settings on their
own and re-parsed the raw document by hand, so the registry could not
guard them and two of them treated a failed load as an empty list.
Approach: loadDesktopSettings() and updateDesktopSettings() (which now
resolves { ok }) replace every direct call; SkillsCatalogPage and
AddCatalogDialog refuse to write the catalog list until it is known.
Testing: bun test packages/ui (403 files), eslint on the changed files.
* refactor(server): validate settings writes against the registry snapshot
Problem: the server whitelist was the only guard on PUT /api/config/settings
and had drifted from the client; dead keys were still persisted.
Approach: settings-helpers.js drops any key the generated registry
snapshot does not list as persistable and strips secret keys from
responses; the dead keys (markdownDisplayMode, toolCallExpansion,
typographySizes, expandedEditorToolbar, gitProviderId/gitModelId) are
gone; the profile keys that were client-only now round-trip. A drift
test requires a valid sample for every persistable registry key.
Testing: vitest run in packages/web (182 files), including the packed
tarball import.
* refactor(vscode): gate bridge settings writes by the registry
Problem: the extension host wrote any key the webview sent straight into
settings.json, and commit-message generation read the dead
gitProviderId/gitModelId pair instead of the small-model setting.
Approach: filterPersistableSettingsChanges applies the registry snapshot
before the file write; chooseBridgeGitGenerationModel honours
smallModelUseDefault/smallModelOverride ahead of the zen fallback.
Testing: bun test packages/vscode (37 files), tsc, build:extension.
* feat(settings): split the user's profile into preferences.json
Problem: one flat settings.json held instance facts, the user's
preferences and device state together, so device state travelled between
installs and the profile had no document of its own to sync from.
Approach: the server keeps one merged document for clients but routes
each key by registry scope on disk (settings-files.js): profile keys go to
preferences.json as { value, updatedAt } entries stamped when the value
changes, everything else stays in settings.json, device keys are dropped
from writes. A missing preferences.json is seeded once from settings.json,
which is left intact; an unreadable one is a failure that pauses profile
writes and never gets overwritten. Server modules that read a profile key
off the disk use the merged sync read. Electron main reads the theme mode
from both files and now owns the splash colours, handed over the
window-theme IPC instead of the settings document. Clients stop sending
device keys, seed them once from a pre-split document, and persist
inputBarOffset locally. The PWA manifest keys are instance facts.
Testing: vitest in packages/web (seed, split write, timestamp retention,
unreadable file), bun test in packages/ui and packages/electron, tsc for
every workspace.
* feat(vscode): write the profile to preferences.json from the extension host
Problem: the extension host writes the shared settings files directly and
had to follow the server's split, and its file writes reported success on
failure.
Approach: settings-files.ts mirrors the server's format and split rules
(seed once, unreadable preferences.json is a failure); persistSettings
routes profile keys to preferences.json and the rest to settings.json,
and the atomic writers now throw so a failed save reaches the webview.
Clearing a key now actually removes it from the owning file.
Testing: bun test packages/vscode (38 files), tsc, build:extension.
* feat(settings): store the per-surface profile fields by surface kind
Problem: theme, chat-layout switches and typography sizes are one value
for every client of an instance, so the phone and the desktop cannot
disagree without a hard-coded runtime branch.
Approach: every settings request carries the client's surface kind in the
x-openchamber-surface header (web, desktop, vscode, mobile — the phone app
and the hosted mobile shell are one kind). For the registry's perSurface
keys the store writes a changed value under fields[key].surfaces[kind] in
preferences.json and never touches the base from a surface; reads resolve
the kind's own value, then the base, then nothing. Writes without a
surface (migrations, the seed) set the base. The VS Code host is always
vscode; Electron main resolves desktop for the native window theme. The
Settings UI is unchanged.
Testing: vitest in packages/web (surface write/read, no base copy, unknown
surface falls back to base), bun test in packages/vscode and packages/ui,
tsc for every workspace, build:extension.
* fix(settings): keep a legacy copy of the profile in settings.json
The first write after the split rewrote settings.json with the instance
part only, and that write happens on startup (relay reconcile). A build
from before the split reads only settings.json, so rolling back would
have lost every preference: theme, default model, all of it.
Every write now stores the profile's base values in settings.json next
to the instance part (`legacySettingsDocumentOf`), on the server and in
the VS Code extension host alike. Current builds ignore the copy because
preferences.json wins in the merged read. When preferences.json is
unreadable the copy already on disk is kept rather than dropped.
Testing: settings-runtime tests updated for the copy; full web suite
(182 files), VS Code tests and extension build, tsc clean. Verified live
on a scratch OPENCHAMBER_DATA_DIR: all 136 keys survive startup, theme
changes land per surface, plain keys land in the base.
* feat(settings): make the UI password and tunnel preset tokens write-only
GET /api/config/settings returned desktopUiPassword and the managed
remote tunnel preset tokens to every authenticated client, including
paired phones and the VS Code webview that never need them.
Both keys are now `secret` in the registry: accepted on write, withheld
from reads. The server answers with a hasDesktopUiPassword flag; the
desktop network page shows "Password set" and sends a value only when
the user types a new one or presses "Remove password" (an empty string
clears it and turns LAN access off). The tunnel page already learned
token presence from the status endpoint. The VS Code bridge strips
secret keys from what it hands the webview while still merging them
from disk on write.
Testing: registry, i18n parity, server settings, VS Code gate tests and
tsc; workspace type-check. Verified against a scratch server: GET
carries the flag and no password, PUT with '' clears, PUT with a value
sets. The desktop-only page itself awaits the owner's run.
* fix(settings): send the surface kind as a query parameter, not a header
The packaged desktop shell (openchamber-ui://app) and the phone app are
cross-origin to the OpenChamber server, so the x-openchamber-surface
header turned every settings request into a CORS preflight the server
did not allow. Settings looked reset and every save reported "Save
failed" without reaching persistSettings. An older remote instance would
refuse the header the same way even with the allow-list fixed.
The client now sends ?surface=<kind>, which keeps the request
CORS-simple on every server version; the server reads the query
parameter and still honours the header. The header is also in the CORS
allow-list for completeness.
Testing: workspace type-check, persistence and registry tests, server
opencode tests. On a scratch server: PUT with ?surface=vscode lands
under surfaces.vscode, GET without or with an unknown surface serves the
base, the header fallback resolves. Confirmed in the owner's rebuilt
desktop and on the phone.
* refactor(settings): drop the show-password toggle from the desktop network page
With the password write-only, the field only ever holds a value the user
is typing right now; the reveal toggle and its strings are gone from
every locale.
* refactor(projects): serve project setup through the server, drop the legacy migration
The shared UI read and wrote ~/.config/openchamber/projects/<id>.json
itself: it resolved the home directory, composed the path, and used the
Files API, which only desktop and VS Code have natively and which cannot
see a remote instance's file at all. It also still carried the months-old
migration from <repo>/.openchamber/openchamber.json, which deleted files in
the folder the upcoming shared project config will use.
The client-owned keys (worktree setup commands, project actions, draft
starters) now live behind GET/PUT /api/projects/:projectId/config.
project-setup.js sanitizes and builds the view; the project-config runtime
merges a patch under the same cross-process lock the scheduled-task writers
hold, so unknown and server-owned keys survive. A wrongly shaped key is a
400, not a silent drop. openchamberConfig.ts keeps its exported functions
and is now an HTTP client. The VS Code webview handles the route locally
and bridges to the extension host, which owns the file with a TS mirror of
the sanitizers.
Testing: server tests for sanitizers, round trip, lock, and invalid patch;
client tests against a mocked route; VS Code sanitizer and bridge tests;
workspace type-check, both VS Code builds, UI isolated suite (409 files),
server projects and project-context suites. Live GET/PUT against a
running server with the owner's real project config.
* feat(projects): read the team's shared config and merge it with the personal one
A project can now carry <repo>/.openchamber/project.json (version 1:
setupWorktree, setupWorktreeWait, projectActions, draftStarters,
plansDir). The server finds the checkout from the path-derived project
id, parses the file, and answers GET /api/projects/:id/config with one
merged view: what runs at the top level, plus shared and personal blocks
so a page can edit the personal file without copying a teammate's entry
into it.
Merge rules: shared setup commands run first (a personal
setupWorktreeMode of "replace" uses the personal list only); the
personal wait flag wins when set; actions union by id with a personal
action replacing the shared one and personal hiddenSharedActionIds
dropping shared ones; starters union by type:name; the primary action is
personal only. A shared file that exists but cannot be parsed, or that
names a plansDir outside the repo, is reported as invalid with a reason
and never treated as "no shared setup". Nothing writes the repo file yet.
Client: getProjectSetup exposes the view; the existing helpers return
effective values, while the Projects page sections and the draft
starters hook edit the personal block only. Shared entries show a quiet
"shared" mark in the actions dropdown and read-only lists above the
editable ones on the Projects page; shared starter chips have no remove
handle. The VS Code extension host mirrors the parser and merge.
Testing: server tests for the parser, plansDir guard, merge table, id
round trip, and a runtime test against a temp checkout; client tests
against a mocked route; VS Code sanitizer, merge, and bridge tests; the
section test covers the shared row; locale parity; workspace type-check;
UI isolated suite (409 files). Live: GET against a temp repo with a
shared file and with a broken one.
* feat(projects): ask before the team's shared commands run, once per set of commands
Shared setup commands and shared actions come from a file a git pull can
change, and they run on the machine of whoever pulls. The first time one
would run, a dialog now shows exactly what would run and asks: "Trust and
run" or "Not this time". A "trust" answer is recorded in the personal
config against a SHA-256 of the executable parts (setup commands and each
action's id, command, and runIn; renames and icons do not count), so a
pull that changes a command brings the prompt back. Nothing asks when the
shared file has nothing that executes.
Worktree creation (session creator, new-worktree dialog, session store,
multi-run launcher, agent-manager empty state) resolves its commands
through the prompt; "not this time" runs only the user's own commands.
The actions dropdown asks before a shared action runs. The Projects page
shows "Trusted on this instance" with a "Reset trust" button next to the
shared actions. The dialog is mounted beside the app-link confirmation on
every shell. The VS Code extension host mirrors the hash and the record.
Testing: server tests for hash stability, ordering, and the trusted flag,
plus a runtime test that changes the shared file and sees trust drop;
client tests for the confirmation store (ask, trust, skip, replace mode,
newer request, failed record, reset); VS Code mirror tests; the actions
button, new-worktree dialog, and issue-2039 tests updated for the trust
path; locale parity; workspace type-check; UI isolated suite (410 files).
* feat(projects): share and unshare setup with the team from the Projects page
The repo file <repo>/.openchamber/project.json is now written by the app,
and only when the user shares something: nothing appears in a repository
until then. PUT /api/projects/:id/config/shared replaces the keys it
names over the current file, writes it pretty-printed with version first
and only the keys that carry something, removes the file (and an empty
.openchamber folder) when nothing is left, refuses a missing checkout or
a plansDir outside the repo, and records trust for the writer, who has
seen what they shared.
On the Projects page, actions and setup commands get "Share with team"
and "Make personal"; shared actions can be hidden for this user; a
checkbox switches to "Use only my setup commands". Project starter chips
get share and make-personal hover buttons. A new "Shared config" block
shows the file's path and status, the shared plans folder, and the trust
status with "Reset trust". A share is a repo write followed by a personal
write; a failure after the first leaves the item visible once, as
personal. The VS Code extension host mirrors the writer.
Testing: server tests for the patch, serialization, emptiness, the write
and removal round trip, the writer's trust record, and the refusals;
client test for the shared route; VS Code bridge test for write and
removal; locale parity; workspace type-check; UI isolated suite (410
files). Live on a scratch server: share, invalid plansDir (400), unshare
to removal of file and folder.
* feat(projects): list, edit, and move plans in the team's shared plans folder
When the shared config names a plansDir, every markdown file in that
folder is a plan on the Plans tab: listed after the user's own plans,
marked shared, addressed as shared:<file>, read and edited in place
(the raw document is written verbatim, so a plan another tool wrote
keeps its shape), and deletable. Share moves one of the user's plans
into the folder; make personal moves it back under a new id; a name
collision gets a numeric suffix. Sharing is refused, with a hint in the
panel, until a shared plans folder is set in Project settings. This
answers the request to read plans from an existing folder such as
docs/plans.
Server: the project-context runtime takes resolveSharedPlansDir from the
project-config runtime; readContext reports sharedPlansDir; POST
.../plans/:id/share and /unshare. Client: movePlan in the context store,
a shared badge and a share / make-personal button per plan row. Session
attachments reference plan ids, so an attached plan that moves has to be
attached again.
Testing: runtime tests for listing, foreign markdown titles, id
traversal, in-place update and delete, share and unshare with a
collision, and the refusal without a folder; HTTP route tests; store and
locale parity tests; workspace type-check; full web suite (183 files);
UI isolated suite (410 files). Live on a scratch server against a temp
repo: list, share, read, unshare.
* fix(server): make OPENCHAMBER_DATA_DIR move every folder, not just the flat files
The variable is documented as the OpenChamber data directory, but only
settings, preferences, auth, and push files followed it; projects,
themes, speech models, and the chats default stayed under
~/.config/openchamber. A second instance started with a custom
directory therefore read and wrote the default instance's project
configs.
Every folder now hangs off the one root. An instance that already used
a custom directory gets projects, themes, and speech-models copied in
once at startup; copied, not moved, so a second instance beside the
default one cannot strip it, and nothing is merged into a folder that
already exists. Existing managed chats are not copied, as with
OPENCHAMBER_CHATS_DIR.
Testing: migration tests for copy-once, no-merge, and same-root no-op;
full web suite; a scratch server with an empty data dir copied the real
project configs and kept its writes in the copy.
* fix(projects): keep a plan's id when it moves into or out of the repository folder
A plan moved into the repository plans folder used to be listed under a
new shared:<file> id, so a session that had attached it lost the
attachment. The manifest entry now stays with a `shared` flag that says
which folder holds the file; the id survives both directions. Only a
plan that never had an entry (one written by another tool) gets an id
when it is brought in. A personal file and a repository file may share
a name because they live in different folders.
Testing: runtime tests for share and unshare with a stable id, reading
and editing the moved plan, the suffix on a name collision, and the
adoption of a foreign file.
* feat(projects): default repository plans folder, "move to repository" wording, tooltips
Plans now have a repository folder without any setup: .openchamber/plans
by default. A custom plansDir replaces the default outright (only that
folder is read and written; moving files between the two is the user's
job), and the field's placeholder and hint say so. The move buttons on
plans are therefore always available.
The word "share" is gone from the UI: it read like publishing, while
the action stores an item in the repository so everyone who pulls it
gets it. Labels are "Move to repository" / "Move to my settings", the
badge is "In repo", the block is "Repository config", and every button
on the Projects page carries a tooltip that says what happens (the
"Move to repository" button explains that edits save first while the
form is dirty). The trust status with "reset trust" moved from the
repository block into the Worktree section next to the commands it
guards; the plan row's badge sits beside the title.
Testing: locale parity, section test, workspace type-check, UI isolated
suite (410 files), full web suite.
* fix(projects): leave the icon key out of the repository file when an action has none
Actions without an icon were written as "icon": null into
.openchamber/project.json. The key is now omitted; readers already fall
back to the play icon. Server and VS Code serializers, tests updated.
* docs: describe the repository config file and how items move into it
A new page in every locale: what stays personal and what can move into
the repository, the .openchamber/project.json format with an example
and every key explained (setup commands, actions with the supported icon
names, starters, plansDir), the merge rules, the trust prompt, and plans
in the repository. Linked from the sidebar and from Project Actions.
Translations written by hand.
|
||
|
Bohdan Triapitsyn
|
85bf0a99de | fix: identify OpenCode Go requests by session | ||
|
Bohdan Triapitsyn
|
6a09c63392 |
feat(small-model): resolve plugin-registered providers from the running OpenCode
Plugin providers are registered from a plugin's `config` hook and credentialed from its `auth` loader, both inside the running OpenCode process. Nothing about them reaches `opencode.json` or `auth.json`, so resolution that only reads files could not see them: selecting such a model failed with "has no known API base URL" while the same model worked in chat (#2666). `GET /provider` is where that state is visible. A new `runtime-providers` module keeps one cached snapshot of it and reports, per provider, the credential and endpoint OpenCode itself resolved. Credential resolution becomes config -> runtime -> auth.json, and endpoint resolution config -> openai default -> runtime -> models.dev catalog. Providers with a dedicated wire format (Copilot, ChatGPT-plan OpenAI, Anthropic, Google) are excluded from the runtime credential: for them OpenCode reports an OAuth access token that their real transport does not accept. opencode zen is excluded when the user has no zen login. OpenCode then reports the sentinel `apiKey: "public"` and trims its catalog to free models that run on its own infrastructure; the sentinel is never read as a credential. Claude Code stays refused for background actions even when a plugin publishes an OpenAI-compatible endpoint for it, because that endpoint is a facade over the Claude Agent SDK and spawns the CLI per request. No capability probe. Asking `GET /models` does identify a plugin whose protocol lives in its own `fetch`, but measured across the 166 providers with an `api` URL in the models.dev catalog it also denies six that work and simply have no `/models` route. A provider that vanishes from the picker explains nothing, while one that fails on use says why, so availability stops at credential and endpoint. The same list drives the Small Model and Changes Walkthrough pickers. Validated against a real OpenCode with four plugin providers loaded: offered providers went from 3 to 7, zen and Claude Code stayed out, and a generation through a plugin-backed model that previously failed now returns. |
||
|
Bohdan Triapitsyn
|
d353638f39 | fix: integrate Claude CLI provider state | ||
|
Cursor AgentandSerhii Dziupin
|
abb396e080 |
fix(walkthrough): block unauthenticated providers with a friendly refusal
When the walkthrough small model resolves to a provider with no usable login, readiness was still ready and generate returned a raw 500 message. Refuse up front with no-provider-login and surface a blocker instead. Closes openchamber/openchamber#2607 Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com> |
||
|
Bohdan Triapitsyn
|
1d17cb87b3 |
feat(walkthrough): write walkthroughs in the reader's language
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. |
||
|
Bohdan Triapitsyn
|
34d0ff7383 |
feat(walkthrough): guided AI walkthrough for diffs, branches, and PRs (#2572)
A diff is ordered by file path, which is almost never the order in which a change makes sense. This adds a Walkthrough surface that reorders it: the model groups related hunks into stops, explains what each group changes about behavior, and orders the stops so each builds on the last. It explains and orders; judging code stays with the existing Review action. Reviews uncommitted work (all, staged, unstaged), a branch against its base, or a pull request. Generation is always user-initiated — nothing runs on a timer, on a file change, or as a side effect of opening a panel. Invariants worth preserving: - Hunk identity is derived on the server and only there. Ids are content hashes, so an anchor that no longer resolves is proof the code it described changed, and staleness needs no heuristics. The client matches ids to ids and never recomputes them; two implementations would have to agree forever. - The digest is never truncated. A diff that does not fit the model's context is refused with an actionable reason, because a walkthrough written against half a diff reads as confident and is wrong. - Nothing disappears. Lockfiles and other generated output are excluded from the model's input by name — never by size — and everything no stop covers is listed at the end, so "have I seen all of it" stays answerable. - Cost is explicit. Results are content-addressed, so returning the working tree to an earlier state costs nothing; generation outlives its request, so a refresh detaches the client rather than discarding paid-for work, and only an explicit cancel stops it. Supporting changes to shared modules: - git: expose the existing getRangeDiff as GET /api/git listUntrackedPaths and getUntrackedDiffs. The latter resolve the repository once for a batch instead of per file, taking a panel ~340ms on an 80-file working tree. - small-model: structured output across four wire forma and abort signal, and an onOverflow policy so an oversized prompt fails loudly instead of being silently clipped. A provider remembered so the prompt-side fallback goes first next time. - models.dev metadata: surface structured_output as tri false blocks a model, a missing field does not, because the catalog omits it for roughly half of all models. Desktop and tablet only: VS Code serves Git through its these routes, and the mobile shell does not consume the surface registry. Docs: packages/docs walkthrough page in English and all eight locales. |
||
|
Andrey Meshkov
|
7e248d4e9b |
Fix: small model dispatch fails for custom OpenAI-compatible providers (#2134) (#2135)
* fix: add support for custom provider base URLs from config * feat: read custom provider apiKey from config and use it as primary credential when it exists * doc: update small-model documentation |
||
|
Bohdan Triapitsyn
|
28f0736d69 |
feat: small-model utility calls on existing OpenCode providers (#2049)
Adds a server-side "small model" capability: direct, cheap LLM calls that reuse the user's existing OpenCode provider logins — the mechanism OpenCode uses internally for titles and summaries but does not expose through the SDK or plugins. Zero new dependencies; plain fetch with per-provider wire formats, credentials never leave the server. Core (packages/web/server/lib/small-model): - Resolution mirrors OpenCode's session scoping: explicit settings override → small_model from the OpenCode config → family scan within the session's provider → the session's own model. The global provider scan only serves callers without a session context, and background callers forbid it entirely (restrictToPreferredProvider), so conversation content never reaches a provider the user didn't pick — explicit choices excepted. - Per-provider auth replicating OpenCode's plugin loaders: GitHub Copilot (device token as bearer, no exchange), ChatGPT plan via the codex Responses API (single-flight OAuth refresh written back to auth.json), Anthropic messages, Google generateContent, generic OpenAI-compatible. - OpenCode's free models (opencode/big-pickle, *-free) are never called directly; unauthenticated providers are skipped by design. - Prompt clamping to the model's catalog context limit; thinking disabled where a wire switch exists (Z.AI/GLM, MiniMax-M3, Gemini Flash); robust content parsing with a clear error when a thinking model spends its whol budget on reasoning. - Settings → Sessions gains a Small Model group: use-default checkbox plus an override picker limited to authenticated providers, persisted with web/desktop/VS Code sanitization parity. Consumers: - Session assist: a server-side watcher on the global SSE hub generates a short recap and one suggested follow-up after a session idles quietly fo a minute, stored on session metadata (openchamber.assist). Freshness is keyed to the last assistant message id, so new activity invalidates the payload everywhere with no extra writes. The chat shows the recap under the last message after five quiet minutes and the suggestion as a dismissible chip above the composer (tap fills the input, never sends). Gated by a new Chat setting (default on) that is a hard generation switch. Language is anchored to the conversation itself, with a script-mismatch guard against model/backend language hallucination. - TTS: a third input mode, summarized — long replies are condensed to spoken prose before playback on any TTS engine. - Git: commit-message and PR generation moved off the active chat session onto the small model fed with real diffs and the commit list (bodies included), with a session-transport fallback for free-model-only setups. - Notes: Add to notes distills long selections into 1-3 dense sentences preserving exact identifiers, with verbatim fallback on failure. Fixes along the way: - The global event watcher now starts unconditionally; it was gated behind the desktop-notify env, leaving the server-side event hub dead in packaged apps. - OpenCode re-emits message.updated for old user messages after idle; the watcher no longer mistakes those for new activity. - Session metadata merges from a fresh read right before the PATCH, so writes made during the generation window (suggestion dismissals, review links) are preserved; the assist runtime stops during graceful shutdown. |