Repository navigation
feat(desktop): add App Dock sidecar tools - #52164
Closed
gustavomhss wants to merge 121 commits into
Closed
gustavomhss wants to merge 121 commits into
gustavomhss wants to merge 121 commits into
Conversation
Adds a Tasks tab to the session side panel (review surface) showing live background work for the current session: subagent cards joined by the task tool's metadata.sessionId (child session id), foreground shell cards, needs-input state from pending permissions, nested (+N) counts, elapsed timers, click-to-open transcript, per-task stop, and auto-open on 0->N. Derives everything from synced reactive stores (no fetching/polling).
Parity test requires every English key present in all locales. Real translations land via the locale sync flow; until then English source stands in (established precedent: 128 identical-EN keys in de.ts).
feat(app): tasks panel for background subagents and shell tools
New optional 'model' param as 'providerID/modelID' (e.g. 'openrouter/deepseek/deepseek-chat'). Resolution: param > agent model > parent model. Provider part selects credentials ambiently (OAuth subscriptions, API keys, custom provider aliases for extra keys). Malformed values fail fast. Docs in task.txt. Tests: forward + reject + unchanged defaults.
When the session carries task rules with model patterns (pattern containing '/'), the resolved provider/model is added to the permission ask patterns and metadata. Otherwise behavior is unchanged (no extra prompt). Denied models surface the session ruleset so the caller can retry with a listed model.
DialogSubagentModels mirrors DialogManageModels (Dialog + searchable grouped List + per-model Switch, live-apply). Reads effective state from session permission rules, appends allow/deny task rules via session.update (server merges, last match wins; first constraint writes the star-slash-star deny baseline). Opened via tasks.models command. English fallback across locales; real translations land via locale sync.
Adds the subagent model picker next to the session model and reasoning selectors: session model, reasoning level (when present), subagent models. New subagentModelsControl slot in session-ui PromptInput; app-side control mirrors the model button styling and opens DialogSubagentModels.
Adds three optional lifecycle hooks to the plugin system, mirroring Claude Code's PreToolUse/PostToolUse/SessionStart/SessionEnd/Stop surface for the cases OpenCode was missing. Events: - session.start: fires when a session is created, before any tool calls. Plugin may tag the session via output.metadata (reserved keys blocked). - stop: fires at end of every agent turn. If the plugin sets output.continue = true and reason === 'completed', the loop re-iterates with the same context. Honored only on 'completed' to prevent infinite loops on aborted or errored turns. - session.end: fires before session deletion, hard-capped at 5s. output.cleanup = false defers forced cleanup within the ceiling. Implementation: - packages/plugin/src/index.ts: adds three optional fields to the Hooks interface (purely additive, no breaking change to existing plugins). - packages/opencode/src/session/session.ts: emits session.start inside createNext; emits session.end inside remove with timeout and error isolation. Plugin.node is wrapped in a lazy Proxy to avoid a TDZ error from the session<->plugin circular import. - packages/opencode/src/session/prompt.ts: emits stop inside the inner 'result === stop' branch where the runLoop is about to break. Honors output.continue by returning 'continue' instead of 'break' so the while loop re-iterates. Tests (6 new, all passing in 4.4s): - packages/opencode/test/plugin/trigger-lifecycle.test.ts: 5 tests using a real plugin file loaded via opencode.json. Verifies that session.start, session.end, and stop receive the documented input shapes; that session.end output.cleanup and stop output.continue mutations propagate; and that stop output defaults to continue=false when the plugin does not touch it. - packages/opencode/test/session/lifecycle-hooks.test.ts: 1 E2E test that creates a session via Session.Service.create and verifies the loaded plugin's session.start hook fires with the correct sessionID, cwd, and timestamp. Docs: - packages/docs/plugins.mdx: full reference for all three hooks (input/output tables, error-handling contract, end-to-end telemetry example). Added to docs.json navigation under a new 'Plugins' group. Error handling: every emit is wrapped in Effect.catch that logs and continues; a throwing plugin never breaks session creation, deletion, or the agent loop. The 5s session.end timeout is enforced via Effect.timeout. Co-authored-by: gustavoschneiter <gustavoschneiter@users.noreply.github.com>
* feat(core): multiple labeled API keys per provider * feat(providers): resolve labeled API keys at runtime in v1 * fix(providers): address cold review findings - resolve getProvider through ensureProvider so keys added post-boot don't race a missing provider entry (crash on first request) - re-read credential state in ensureProvider to pick up remove/update - fail loudly in v2 runner when a virtual credential is missing instead of silently routing to a different active key - respect provider.disabled for virtual provider availability - redact API keys from the public provider list response - honor disabled/enabled provider sets in the post-boot catalog fallback - route vars/model loaders to the base provider for virtual ids * fix(app): persist labeled keys via v2 credential route connect.key was overridden to write legacy auth.json, dropping the label and never creating a credential row. Try the v2 integration connect.key (which stores a labeled credential) first, falling back to the legacy auth write only when the server lacks the v2 route. * test(app): cover v2-first connect.key with legacy fallback * fix(providers): address second cold review round - server-compat connect.key falls back to legacy auth only on missing-route (404/405), re-throwing real credential failures (400/401/500) instead of silently writing legacy auth on top of a broken key - getProvider now fails typed with NoProvidersError; update the existing provider test that asserted the old undefined-return, and hoist the getProvider error channel in the Service interface so consumers handle the typed failure - fix parseVirtualID usage in vars/model loader lookup (baseID, not base) and type the credential id lookup --------- Co-authored-by: Gustavo Schneiter <gmhschn@gmail.com>
Composer button and dialog now work on drafts: toggles accumulate in a draft-scoped pending set, flushed into session permission rules right after session creation and before the first prompt is sent (sequential, no race). Button visible in both V2 and legacy composers; sessionID threaded via props.
…tion describeTask appends the session's allowed provider/model patterns (same last-match-wins semantics as enforcement) so the model sees exactly what it may pass as task model, plus a pointer to the parameter. No session rules means no section (unchanged behavior).
Child cards are now keyed exclusively by child session id (children() authoritative, task tool parts enrich or fall back to orphans) — fixes duplicate agent cards. childStats reads the promise-client message shape (type: assistant, model: {id, providerID}) alongside the legacy flat fields — restores per-card model + tokens. Headlines strip the ' (@agent subagent)' suffix.
feat: subagent model allowlist + tasks panel polish
Co-authored-by: Gustavo Schneiter <gmhschn@gmail.com>
* refactor(opencode): lazy-load CLI commands to cut startup cost The entrypoint statically imported every command module (server, session, sdk, config schemas) at process start, costing tens of seconds and ~200MB RSS even for fast paths like --version. Register commands through a delegating module that only loads the real implementation when yargs parses an invocation that exercises it. Measured: --version instructions 10.4B -> 5.6B (-46%), peak RSS 224MB -> 132MB (-41%), user CPU 3.4s -> 1.9s (-44%). Also decouple UI.cancelled-error from the effect import so the ui module loads without pulling in the full effect runtime. * perf(core): skip durable event log for message/part updates in local mode updateMessage/updatePart published full snapshot events to the durable event table on every change, duplicating the complete payload per update. A single 23MB message re-rendered 90x wrote ~2GB of redundant JSON to the event log (the user's opencode.db grew to 7GB, 6.4GB of it the event table). The only consumer of these durable rows is the experimental workspaces sync (off by default); the local UI/SSE/projection read the projected message/part tables. Add PublishOptions.persist (default true = unchanged). updateMessage/ updatePart pass persist: experimentalWorkspaces so local installs stop appending to the event log while still projecting and notifying. With workspaces enabled the behavior is byte-identical to before. The bridge only emits sync envelopes for events carrying a durable marker, so local events are never observed by cross-instance sync. Measured: write-path E2E (90x growing part up to 24MB) 10.2s -> 5.5s (-46%); publish micro-bench -92%; durable bytes per turn 100% eliminated. Add db compact command as an optional one-shot tool to reclaim rows already written (guarded against experimentalWorkspaces, which requires contiguous seq for sync replay). * refactor(opencode): type lazyCommand without any * test(core): drop timing-based persist benchmark The publish timing bench was flaky under machine load and duplicated the behavioral coverage already provided by event-persist-gate. The measured speedup (write-path E2E -46%, publish micro-bench -92%) is captured in the perf commit message instead of as a timing assertion. * docs(core): explain persist:false seq placeholder and commit hook contract Reviewers need the invariant spelled out: the local-only publish path supplies an inert seq placeholder to projectors (none read it) and skips commit hooks by design (no caller combines them). Adds the rationale for gating message/part persistence on experimentalWorkspaces in session.ts. * feat(opencode): add native HuGR composer tools --------- Co-authored-by: Gustavo Schneiter <gmhschn@gmail.com>
Adds 'opencode debug read --params <json>' to invoke the legacy read tool directly without an LLM. Emits a machine-readable envelope: params, latency ms, title, metadata (incl. telemetry source/size/lines_read/saved) and output. Failures come back as ok:false with the error. Runs under Effect.scoped so LSP deps resolve.
bench-read.sh runs N invocations of 'opencode debug read' across whole/symbol/tail modes against a target file and emits CSV (mode,runs,avg,p50,p95,size,lines_read,saved,source). Supports BENCH_SYMBOL override and --json for full envelopes. Bash-only (no %f printf; stats via python).
…ark harness findSymbolByIndent and tail line-count now share a path+mtime+size memo scan cache, fixing two defects the benchmark exposed: (1) symbols past line 2000 fell back to an unbounded scan every call even warm; (2) tail's negative-offset line count re-read the full file uncached (~1s on 15k-line files). Scan cache is bounded at 200k lines. debug read now accepts a params ARRAY (warm in-process runs), --meta-only (envelope without output payload for benchmark-safe huge reads), and emits envelopes with write-flush. Ship bench-corpus.sh (3 sizes x 2 densities), bench-hardcore.sh (cold+warm matrix over whole/symbol/tail/depth), bench-report.py (latency, output-bytes deltas, resolution source, slice economics).
…view Two hardening deltas from the cross-market comparison (Claude/Codex/Grok), benchmark-validated against the corpus: 1. Unresolvable symbol now DEGRADES to a normal file read with a <system-note> instead of failing — the model never pays a second round-trip on a probe name. Symbol resolution contract unchanged on the happy path (LSP → indentation → render); failure path switched from hard error to graceful continuation. 2. Token budget (25k est, ~2 bytes/token conservative bound) cuts dense content earlier than the 50KB byte cap with a 'PARTIAL view' footer + next offset, so token-dense reads never dump a context bomb (Grok parity). Code ASCII fires the gate co-incident with the byte cap; tests updated: symbol-not-found now asserts degradation; new PARTIAL-view truncation test. Benchmark (v2, same corpus): no regression — symbol slice economics identical (15,005/15,009 lines saved large-dense; -97..-99% output), resolution accuracy unchanged, cold→warm unchanged.
…g + metaprompt Extends the read tool with the async-resolve capabilities recommended after the cross-market comparison + RL adoption research: - search='<fragment>' — find a symbol by name fragment and read ONLY that range, auto-resolving via LSP → indentation heuristic → degrade, so the model can request a precise slice without knowing the exact symbol name and without paying a second round-trip on a miss. Shares the resolveRange path with symbol; source reported. - expand: sparse=true numbering (anchor 1 + every 10th, Grok-style) opt-in to cut numbering tokens on huge reads; default unchanged. - read.txt metaprompt: instructs models to prefer symbol/search over whole-file reads when a name/fragment is known, since the tool degrades safely on a miss — nudging RILI(RL) adoption without changing the contract or adding decision friction. Tests 45/45 (new: search-by-fragment, sparse anchors). Benchmark v3: search delivers -99% output on dense files (58,470 → 469 chars), consistent with symbol; no regression on whole/tail.
added 22 commits
September 9, 2026 16:35
# Conflicts: # packages/opencode/src/session/compaction.ts
Remove status===running guard in PtyHttpApi.get handler so exited PTY sessions with their exit code are returned by the API. This enables the PTY test to observe exit 4 → status: exited, exitCode: 4. Co-authored-by: opencode
- Add app-dock-rpc.ts: bridge between sidecar and Electron main process for dock_* ops - Add app-dock-browser.ts: injected scripts (snapshot, click, type, scroll, navigate, etc.) - Add app-dock-api.ts: TypeScript interface for AppDock - Add app-dock-tools.test.ts, app-dock-rpc.test.ts, app-dock-youtube.test.ts, app-dock-live.test.ts - Extend app-dock.ts with execute(), scroll(), hover(), drag(), clickAt(), scrollTo(), storage(), evaluate(), network(), pdf(), list(), closeAll() - Register RPC bridge in ipc.ts with profile resolver and new IPC handlers (app-dock-list, etc.) - Update server.ts to accept sidecarPath and forward messages for RPC - Add preload API: appDockList, tab-selected event - Add opencode plugin: app-dock.ts with 16 dock_* tools (open, read, click, type, navigate, go, list, close, scroll, wait, screenshot, keyboard, evaluate, storage, network, activate) - Add corpus docs: invariants, requirements, goldens, cards for app-dock
Add AppDockPlugin to internalPlugins when flags.client === 'desktop' so the sidecar loads dock_* tools when running in desktop context.
…Script, buildRetryScript) - Fix typo: buildPDFSript -> buildPDFScript - Fix buildFrameScript: cross-origin iframe access handling - Fix buildRetryScript: use string concat to avoid TS parser bug with template literals - Update CI workflow to remove non-existent app-dock-test-support.ts
- Convert module singletons to AppDockRPC class with explicit init/reset - Add resetAppDockRPC() export for test isolation - Fix handleDockRPC to use proper promise chain (no async IIFE leak) - Remove debug logging from app-dock-tools.test.ts - Add 60s watchdog in child process to prevent hangs
- Add retry logic for read after reload to handle namespace stabilization - Increase wait times after reload (500ms) and before fullscreen read (300ms) - Window focus is now more reliably handled with registerAppDockWindow() call
github-actions
Bot
added
the
needs:compliance
Contributor
|
This PR doesn't fully meet our contributing guidelines and PR template. What needs to be fixed:
Please edit this PR description to address the above within 2 hours, or it will be automatically closed. If you believe this was flagged incorrectly, please let a maintainer know. |
Contributor
|
The following comment was made by an LLM, it may be inaccurate: |
gustavomhss
marked this pull request as ready for review
September 29, 2026 20:26
Author
|
Opened against wrong upstream repository; closing. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Context
Adds desktop-only App Dock tools for sidecar LLM browser control.
What
Tested
Risk / Rollback
Large feature diff: 63 files, including generated App Dock corpus docs. Draft pending split/review decision. Revert commits 941082b through afdc50d to roll back.