feat: re-queue remaining chain steps on global-deadline expiry#70
Open
JeanBaptisteRenard wants to merge 89 commits into
Open
feat: re-queue remaining chain steps on global-deadline expiry#70JeanBaptisteRenard wants to merge 89 commits into
JeanBaptisteRenard wants to merge 89 commits into
Conversation
Subagent transcripts written by Claude CLI live at
<folder>/<parentSessionId>/subagents/agent-<agentId>.jsonl alongside a
.meta.json sidecar holding { agentType, description }. Surface them as
first-class rows in session_cache, keyed by sessionId 'sub:<parent>:<agentId>'.
- adds parentSessionId, agentId, subagentType, description columns
- migration v4 clears the cache so a re-index repopulates everything
- adds idx_session_cache_parent for hierarchy lookups
- new query getCachedByParent + widened cacheGetByFolder
Adds helpers for the on-disk subagent layout: - enumerateSessionFiles(folderPath) walks top-level + <uuid>/subagents/*.jsonl + legacy <uuid>/*.jsonl fallback - subagentSessionId(parent, agentId) gives the synthetic 'sub:<p>:<a>' id - resolveJsonlPath(projectsDir, row) reconstructs the absolute path - readSubagentMeta reads the .meta.json sidecar readSessionFile now accepts opts.parentSessionId. When set, it requires isSidechain on at least one entry (defensive guard), reads the sidecar for agentType/description, falls back to filename for agentId if absent in entries. Adds 10 unit tests covering both branches, the sidecar guard, the legacy fallback path, and the synthetic-id format.
Both scanners (synchronous refreshFolder + worker scan-projects) now use enumerateSessionFiles instead of a flat top-level readdir, so subagent transcripts get cached as first-class rows with parentSessionId set. FTS indexing inherits this for free: subagent summaries land in search_fts the same way top-level sessions do, with type='session' and folder=<parent>. Full-text search now finds anything an Agent call discussed.
Clicking an Agent tool_use block in the JSONL viewer now loads the matching subagent transcript and renders it nested below the block. Re-click collapses without refetching. - main.js: new IPC read-subagent-jsonl(parentSessionId, agentId) and list-subagents(parentSessionId) - preload.js: window.api.readSubagentJsonl / listSubagents - jsonl-viewer.js: clickable Agent block, caret indicator, recursive nested render. Identical (description, subagentType) pairs are disambiguated by occurrence ordinal so parallel fanout calls each open their own transcript. - style.css: subtle hover + left-border indent for nested transcripts
Runs `npm ci && npm test` on ubuntu-latest against Node 20 and 22. Triggers on pull requests targeting main and on direct pushes to main.
Migrations that clear session_cache (v2, v3, v4) leave the cache empty on first launch. The previous get-projects handler returned [] immediately and fire-and-forgot the worker scan, relying on a later 'projects-changed' IPC to trigger a reload — but app.js only reloads on that event when the user happens to be on the Sessions tab. If they were on another tab (or hadn't navigated to Sessions yet), the list stayed empty until they manually switched tabs. Make populateCacheViaWorker return a Promise that resolves when the scan finishes. Concurrent callers share the same Promise. get-projects awaits that Promise when the cache is empty, so the response carries the freshly populated cache and the renderer paints immediately.
buildProjectsFromCache was projecting only the pre-PR#1 columns onto the session object sent to the renderer. The renderer's hierarchical sidebar (introduced in PR#2) reads parentSessionId/agentId/subagentType/description from each session and silently flattens to top-level when they're missing — which is what happened: all subagents appeared as siblings of their parent instead of nested under it. Add the four subagent columns to the projection. No DB or wire change.
Subagent support: index, search, and inline-expand transcripts (fork CI)
readSessionFile JSON.parse'd every line of a session file in a single outer try/catch. If a Claude CLI session was actively writing the file while the worker scan read it, one mid-write line could throw and invalidate the ENTIRE file's session row. With many parallel live sessions this manifested as 'most projects show no sessions after a fresh index'. - Per-line try/catch inside readSessionFile: skip malformed lines, keep parsing the rest. - Per-file try/catch in workers/scan-projects.js: defensive belt and braces so one unparseable file can't abort an entire folder scan.
resolveWorktreePath was defined and exported in the file but never actually called from deriveProjectPath. As a result every worktree under <project>/.claude/worktrees/<name>/ appeared as a separate project group in the sidebar instead of being grouped under its parent project. Wire the call in both branches of deriveProjectPath (direct .jsonl path and subdirectory path) and export resolveWorktreePath for other callers.
session-transitions.js now tracks per-active-session knownSubagents and emits two new IPC events: - subagent-spawned (parentSessionId, agentId, subagentType, description) when a new agent-*.jsonl file appears under <sid>/subagents/ - subagent-completed (parentSessionId, agentId) when an existing file's mtime has been stable for >30s Adds two IPCs for read-only live tailing: - start-subagent-watch (parent, agentId) -> watchId - stop-subagent-watch (watchId) The watcher streams new JSONL entries via subagent-watch-event so the renderer can append them to an open inline-expanded subagent transcript without polling.
- sidebar.js: subagents are no longer flat siblings of their parent. Each top-level row now shows a 'N subagents' affordance with a disclosure caret; expanded children render indented with a subagentType pill, description as label, and persist expansion state per parent in localStorage. Orphan subagents (parent absent from cache) hoist to a 'Orphan subagents' group. - grid-view.js: each active session card now shows a stack of small pills for currently-running subagents, color-coded by subagentType, capped at 5 with a '+N more' overflow. Listens on onSubagentSpawned/onSubagentCompleted. - jsonl-viewer.js: when an inline-expanded subagent block represents a still- running agent, start an fs.watch via the new IPC and append streamed entries; show a small '● live' indicator until completion. Stops the watch on collapse or subagent-completed. - style.css: minimal styles for new sidebar/grid/live elements, matching existing palette and font weights.
When Switchboard attaches to a session that already has many subagent files on disk (e.g. a long-running session with 100+ Agent calls), the first walk of detectSubagentTransitions treated every existing file as a 'first sighting' and emitted subagent-spawned for each. 30s later it emitted subagent-completed for each. Hundreds of IPC events back to back froze the renderer UI. Distinguish bootstrap (first walk for this session) from steady-state. On bootstrap, record every existing file in knownSubagents silently: - files modified in the last 60s stay in the active lifecycle (could be mid-run, will eventually fire subagent-completed) - older files are marked completed immediately with no IPC Only files that appear AFTER the bootstrap walk fire subagent-spawned.
…lity Subagent observability: hierarchy, transitions, badges, live tail (fork CI)
…om disk
The existing 'Hide worktree' affordance only added the path to hiddenProjects;
on-disk worktrees accumulated. This adds a real delete action.
- main.js: new IPC delete-worktree(path). Validates the path matches the
worktree-layout regex (.claude/worktrees, .claude-worktrees, .worktrees),
runs 'git worktree remove -f' via execFile (no shell), retries with '-f -f'
on locked worktrees, then purges any matching session_cache rows and FTS
entries. Returns { ok, removed } / { ok: false, error }.
- preload.js: window.api.deleteWorktree binding.
- sidebar.js: 'Delete worktree' button on worktree sub-groups, with confirm
dialog. Keeps 'Hide worktree' for the cosmetic case.
Hide is unchanged; Delete is opt-in and confirmed before running.
Makes it possible to tell the dev build apart from a released installed binary in the OS task switcher, dock, About menu, and window title. No-op in packaged builds.
Adds 14 unit tests for two load-bearing behaviours shipped via PR#1/#2 that previously had no coverage: - derive-project-path.js — resolveWorktreePath collapses worktree paths (.worktrees, .claude-worktrees, .claude/worktrees) onto the parent project; deriveProjectPath end-to-end via stubbed jsonl - session-transitions.js — bootstrap silent-init (no spawn/complete events on first walk), post-bootstrap spawn detection, completion timing
Worktree cleanup UI (fork CI)
…ons-worktree test: coverage for resolveWorktreePath + subagent cold-start
refreshFolder used a linear scan of cachedMap for every on-disk file (O(N²) on a folder with N cached sessions). For projects with thousands of subagent transcripts this froze the main process on every fs.watch flush — fs.watch fires often while live Claude sessions append JSONL, so the freeze recurred every 500ms (the debounce interval). Build an inverted filePath → dbId index once and look up O(1) per file. Confirmed: 24/24 tests still pass.
Even with the O(1) lookup, refreshFolder still ran enumerateSessionFiles
(many readdirSyncs) and fs.statSync on every file in the folder on
every flush. For projects with thousands of cached subagents and many
concurrent active agents writing JSONL, the main process kept blocking
on syscalls every 500ms.
The watcher already knows which file changed. Plumb that information
through to refreshFolder via opts.files; in targeted mode, skip
enumerateSessionFiles entirely and only stat the dirty paths.
- main.js: pendingFolders (Set) → pendingChanges (Map<folder, Set<relPath>|true>)
- session-cache.js: refreshFolder(folder, { files }) — when files is a
non-empty Set, scan only those entries; otherwise full walk
- Targeted-mode deletion: rely on per-file ENOENT in statSync, skip the
whole-folder GC sweep (no longer accurate when we only saw a subset)
- Falls back to full walk for folder-level events / bootstrap
24/24 tests still pass.
Live Claude session JSONLs grow without bound — observed a 218 MB host-session file in a real workload. refreshFolder ran fs.readFileSync + JSON.parse on the full file every time the watcher fired (every ~500 ms while the session was being written), making the main process freeze for 1-2 seconds at a time. When a file is already cached and now exceeds HUGE_FILE_BYTES (5 MB), skip the re-read entirely: just bump the modified timestamp in the DB so the sidebar shows activity. Summary/slug/title were captured when the file was smaller and rarely change after the first turn; the next cold start or a shrink below threshold refreshes them. Adds db.touchCachedModified(sessionId, modified) — a one-row UPDATE that avoids the 14-column upsert when only mtime matters. 24/24 tests still pass.
…w ones User feedback: "for display we shouldn't need to go that deep into the file — full read should only happen for search". Implements that split. readSessionDisplayHeader: stream-reads the first ~256 KB / 500 lines and extracts only what the sidebar needs (summary, slug, customTitle, aiTitle, agentId, isSidechain marker, subagent sidecar). No textContent, no messageCount. ~1 ms even for a 200 MB host-session JSONL. refreshFolder flow: - NEW file (no cache row) → full readSessionFile, seeds FTS body - EXISTING file → header-only refresh, merges fresh display fields with cached body. NO FTS write — search index for live sessions lags until cold-start - Header fails → mtime-only touch as last resort cacheGetByFolder widened to SELECT * so refresh can merge unchanged fields (created, messageCount, textContent) without re-reading. Drops the HUGE_FILE_BYTES hack from the previous commit — the header approach handles size uniformly so no special-casing. 24/24 tests still pass.
The header-only refresh leaves search_fts stale for active sessions —
content typed after the last cold-start isn't indexed. Adds a user
trigger that runs the full worker re-scan (which rewrites FTS from
the live JSONL tails) and then re-fires the current query.
UI
- New refresh button inside the search bar (circular-arrow icon)
- Enter in the search input triggers the same path
- Spinner on the button while reindexing
- Search debounce bumped from 200ms to 350ms (gentler under heavy
concurrent workloads)
Wiring
- main.js: ipcMain.handle('rebuild-cache') → populateCacheViaWorker
- preload.js: window.api.rebuildCache()
- app.js: runSearchQuery() extracted; triggerRebuildAndSearch()
serialises rebuild + refire and guards against double-clicks
24/24 tests still pass.
Renderer threw ReferenceError on every project iteration because the result destructure at sidebar.js:438 was missing 'subagentIndex' even though processProjectSessions returns it and buildSessionsList expects it at line 477. The error aborted the project loop, leaving the sidebar blank while the backend correctly returned 13 projects / 1500+ sessions. Pre-existing bug from PR#2's hierarchical sidebar — became visible now because every project in this workload has subagents. Also nudges #search-refresh-btn from right:42px to right:60px so it no longer overlaps with #search-clear (the × button at right:40px).
…efault On long-lived projects the orphan-subagent list can grow huge (>1000 in this session's host project) and pushes the rest of the project out of view. Default the section to collapsed and let the user toggle it with a click on the label. Adds a right-pointing caret that rotates 90° when expanded and a per-project state in localStorage so the choice sticks across reloads. - localStorage key: 'orphanExpanded:<projectPath>' = '0' | '1' - Default: collapsed (no key set) - Label format: '▸ Orphan subagents <count>'
The previous commit referenced 'project.projectPath' inside buildSessionsList, which never had 'project' in scope — runtime ReferenceError on every render, sidebar blank again. Pass projectPath as an argument from both call sites (regular projects and worktrees). Also leaves the renderer console→main bridge in place under mainWindow.webContents 'console-message' — paid for itself twice now.
Live JSONL writes fire the watcher every ~500ms, triggering notifyRendererProjectsChanged on each flush. Even with the header-only refresh, the renderer was re-fetching projects and running morphdom diff over 100+ session items at that cadence, producing visible sidebar flicker. User flagged it as 'UI glitch on left side during refresh'. - session-cache.js: leading-edge throttle on notifyRendererProjectsChanged, 1.5s cooldown with trailing flush so the first change is instant but bursts coalesce. - public/app.js: bump renderer debounce 300ms → 900ms. Combined with the main-side throttle the sidebar redraws at most ~1×/sec under heavy load.
…ookup perf(refresh): O(1) cached lookup in refreshFolder
Wraps npm scripts under task(1) with named tasks (install, dev, build, test, lint, check, ci, clean, db:reset, install:lint). Updates README with a Tooling section documenting the preferred workflow.
* fix(app): showSession() on openTerminal error path * feat(sidebar): detect missing project paths * feat(main): remap-project IPC with atomic JSONL rewrite * fix(remap-project): address CRITICAL review findings - CRITICAL-1: replace flat readdirSync with enumerateSessionFiles so subagent transcripts under <uuid>/subagents/*.jsonl and the legacy <uuid>/*.jsonl layout are also rewritten - CRITICAL-2: add active-sessions guard — refuse remap if any non-exited PTY session has projectPath matching the folder (avoids concurrent-writer data loss); also re-check fs.existsSync(oldPath) at handler entry so a path that came back does not get clobbered - Extract rewriteJsonlAtomic helper that cleans up orphan .tmp files on error (try/catch around writeFileSync + renameSync) - Switch statSync → lstatSync on newPath to make symlink intent explicit - Tests: add subagent layout (preferred + legacy), orphan .tmp cleanup, and active-sessions guard tests (4 new tests, total 51 passing / 64)
* docs: AI agent guidelines + AppImage replacement safety CLAUDE.md follows the workspace convention (single-line @ pointer) with content in .ai/shared-guidelines.md. Covers: - Invariants for AI agents working in this fork (no double Electron, SWITCHBOARD_DATA_DIR isolation, worktree isolation for parallel agents, no Co-Authored-By). - Fork-specific features not in upstream (subagent support, work-files tab, heatmap from DB, single-instance-lock, etc.). - Helpers worth reusing (enumerateSessionFiles, encodeProjectPath, ViewerPanel, escapeHtml). - Upstreaming workflow. README.md gets a "Working alongside a running AppImage" subsection that explains why rebuilds + cp don't kill the running instance (the live process extracted /tmp/.mount_* at launch and doesn't need the on-disk file anymore). * chore(ai): complete init-project scaffold - .ai/shared-guidelines.md: import @~/.skaleet-ai/conventions/rules.md for universal Skaleet AI rules (HANDOFF, model gate, worktree isolation, no Co-Authored-By, shell-command pitfalls, memory hygiene). Caveat block notes which rules don't apply here (DDD/CQRS, docker compose exec, monitor-ci, glab, /pre-commit — Switchboard is Electron, not a Skaleet backend service). - .ai/project.json: scopes (main / renderer / tests / build) with verify commands so /dispatch knows what to run. Encodes the no-double-electron and worktree-required constraints. - Memory dir initialised at ~/.claude/projects/-home-jean-baptiste-workspace-switchboard/memory/.
Add 5 focused context docs at .ai/contexts/*.md covering the natural
boundaries inside Switchboard:
- session-cache: SQLite + FTS5 + watcher + targeted refresh
- schedule-runner: in-process cron + .md storage + claude --resume -p
- subagent-observability: parent->child grouping + transcript view
- viewer-panel: reusable CodeMirror panel for Plans / Memory / .work-files
- ipc-bridge: trust boundary inventory (every IPC, every event)
Each doc: ~70-130 lines, structured around purpose, key files, public
surface, invariants, non-obvious behaviors, and "if you change this,
also check..." cross-references. Audience is AI agents who need to make
a focused change without re-reading 1800 LOC of main.js.
Also adds:
- .ai/contexts/README.md index + recommended reading order
- .ai/contexts/_issues.md observations captured while writing (not
blockers, opportunistic follow-ups)
- .ai/project.json contexts map
- shared-guidelines.md quick-orientation table now points at the
context docs instead of raw files
Follow-up to the context-engineering skill refactor: - .ai/project.json gains "context_engineering_profile": "electron-flat" so the new router skill detects the profile without falling back to heuristics. - AGENTS.md is a thin pointer at the repo root referencing the universal rules and .ai/shared-guidelines.md. Conventional location agents look at first; mirrors the CLAUDE.md pointer pattern.
…scripts (#24) * feat(main): trigger-watcher — file-based input injection for harness scripts Drop a JSON trigger file into ~/.switchboard/triggers/<uuid>.json to send a command (e.g. /compact) into any open PTY session. Optional idle-wait polls session._cliBusy (the OSC-title spinner flag) before writing, so the harness can safely call /compact without racing mid-response output. - trigger-watcher.js: watches triggers dir, processes JSON files, writes result to processed/<uuid>.result.json, deletes the trigger on done. - main.js: wires start() after startScheduler with getPtyForSession and isSessionBusy closures over activeSessions. - test/trigger-watcher.test.js: 6 node:test cases (happy path, unknown session, malformed JSON, missing field, idle-wait flip, idle timeout). - .ai/contexts/trigger-watcher.md: context doc. - .ai/contexts/README.md: table entry. * fix(trigger-watcher): address code review findings - C1: lstatSync size guard — reject trigger files > 64 KB before readFileSync - C2: lstatSync isFile() — reject symlinks and non-regular files - W1: SyntaxError retry — wait 50 ms and re-read once before failing on partial writes - W2: command length cap — reject commands > 4 KB - W3: control char sanitization — reject \r \n \0 \x1b in command - W4: concurrency cap — queue triggers when > 8 in-flight; simple semaphore via waitQueue - W5: session-exit during wait:idle — check getPtyForSession on each poll tick - W6: tests for C1, C2, W1 retry-then-fail, W2, W3, W4 (12 concurrent), W5, PTY throw, dedup, I4 - I1: atomic result write — write to .tmp then renameSync to final path - I3: wrap trigger-watcher start() in try/catch in main.js boot path - I4: NaN guard on parseInt(SWITCHBOARD_TRIGGER_IDLE_TIMEOUT_MS) — fall back to default
…#25) The "Building does NOT kill the running instance" rule (added in the docs/ai-guidelines PR) was wrong. Witnessed 2026-05-31 ~13:49 — running AppImage went silent in main.log during a background `npm run build:linux`, no SIGTERM logged, 9-minute gap before user relaunched. Suspected cause: electron-builder's @electron/rebuild replaces native `.node` files (better-sqlite3, node-pty) via non-atomic truncate+write, breaking the dlopen handle in the running process → SIGKILL. The `cp dist/*.AppImage ~/Applications/...` step itself is safe — verified at 17:36 (process kept running through the swap). Updates: - .ai/shared-guidelines.md §2 — rewritten as "build is risky, cp is safe" - README.md "Working alongside a running AppImage" — same correction - .ai/contexts/_issues.md — full incident writeup + mitigation candidates
…rigger timeout_ms Raises DEFAULT_IDLE_TIMEOUT 30s→5min, adds MAX_TRIGGER_TIMEOUT=600_000ms cap, adds optional timeout_ms per trigger field. Precedence: per-trigger > env var > default. Validated end-to-end by two production self-compacts (2026-05-31). 86 tests pass. Code review APPROVE (3 non-blocking NITs deferred).
node-pty's ptyProcess.write() is silent on a dead child. Adds defaultIsPtyAlive() using process.kill(pid, 0) signal-0 probe, with checks at both lookup time and right before write. 3 new W7 tests (25 total pass). Critical for safe deployment of auto-compact harness in routine.
Adds a chain field to the trigger JSON for sequential multi-step PTY injection. Mutually exclusive with command. Each step waits for the previous turn to complete (busy-rise + 2s TOCTOU window, then busy-fall) before injecting the next. Global and per-step timeout_ms supported, per-step capped at remaining global. W7 liveness checks applied at chain entry AND before each step's write (defense-in-depth). 37 tests pass (22 existing + 11 chain CHAIN-1..11 + 3 W7 + 1 CHAIN-12 covering the instant-reply path). Final review APPROVE.
The stats screen renderer was already wired to display tokens, tool calls, and per-model usage, but the backend only fed messageCount — so those cards showed zeros and Total Sessions was inflated by subagents. - read-session-file.js: extractDailyMetrics() accumulates per-(date,model) tokens (input/output/cache), tool_use counts, and message counts bucketed by each line's timestamp. Synthetic/model-less assistant turns bucket under model '' with no tokens; tool_result-only user turns are not counted as messages. dailyMetrics attached to parent + subagent return objects. - db.js: migration v5 adds session_metrics(sessionId,date,model,...) + date index; replaceSessionMetrics() (delete-by-session then insert, in a txn); deletion wired into deleteCachedSession/deleteCachedFolder (folder delete sub-selects on session_cache, so it runs first); aggregates getDailyMetrics/getDailyModelTokens/getModelUsage/getTotalCounts (totalSessions = parents only). - session-cache.js: writes metrics on the full-read paths only (NEW-file branch of refreshFolder + the worker handler), not the header-only refresh. - main.js: get-stats-from-db / refresh-stats now build the full stats object (dailyActivity, dailyModelTokens, modelUsage, totals) from session_metrics. - public/stats-view.js: two new summary cards (Total Tokens, Tool Calls). Tests: test/read-session-file-metrics.test.js (extractDailyMetrics unit + readSessionFile integration), test/db-session-metrics.test.js (pure-JS mirror of the SQL aggregation). Metrics populate automatically on next cold-start rebuild — no separate backfill.
Two coupled changes:
1. chain field — a trigger may carry `chain: [{command, timeout_ms?}, ...]`
instead of a single `command`. Steps are injected sequentially, each
waiting for the prior turn to complete (busy then idle) before the next,
under a shared global deadline. Lets one trigger drive a /compact then a
resume prompt without racing on a shared idle tick.
2. discrete-Enter submit — write the command text and the Enter keypress as
SEPARATE PTY writes (SUBMIT_ENTER_DELAY_MS apart). A CR concatenated onto
the text in a single write is absorbed by Claude Code (kitty keyboard
protocol) as a literal newline and the command never submits; only a
discrete CR submits, mirroring how xterm sends each keypress. Fixes
free-text trigger commands landing in the composer unsubmitted while the
short menu-driven /compact path appeared to work.
Tests: 101 passing (node:test). Write-shape assertions updated for the
two-write submit; makeChainCtx starts a turn only on the discrete Enter.
Call populateCacheViaWorker() unconditionally in app.whenReady so deleted transcripts (sub-agent / workflow runs cleaned up between sessions) are evicted from session_cache on the next launch. The worker does a deleteCachedFolder + upsert per folder — a full prune — and runs in a Worker thread so startup is non-blocking. Concurrent callers (FTS-recreated path, first get-projects call) share the in-flight Promise via the existing guard, so no double scan occurs. Previously the prune only ran when isCachePopulated() was false (cold DB after migration) or when searchFtsRecreated was set, leaving ~457 stale rows visible in the sidebar across normal restarts.
…screte-submit fix(trigger-watcher): discrete-Enter submit + chain triggers; prune stale sidebar cache
…ection feat(stats): collect token/tool/message metrics for the stats screen
…sionCache ctx.db
main.js hand-builds the db: {} allow-list passed to sessionCache.init(). The
stats work added replaceSessionMetrics to db.js + session-cache.js but never
added it to that literal, so ctx.db.replaceSessionMetrics was undefined and the
call inside the cold-start worker handler threw on stderr (not electron-log),
silently aborting the indexing write loop → session_metrics stayed empty → the
stats screen still showed 0 tokens / 0 tool calls after the build.
touchCachedModified was missing from the same literal (pre-existing latent bug,
header-refresh fallback path) — added too.
Adds test/main-ctx-db-wiring.test.js: static guard asserting main.js's ctx.db
allow-list covers every ctx.db.* session-cache.js dereferences. Verified it
flags both names against the pre-fix main.js. No unit test booted this wiring
before, which is why task check stayed green through the bug.
fix(stats): wire replaceSessionMetrics + touchCachedModified into sessionCache ctx.db
The main process pegged ~53% CPU continuously while idle. Three causes: - main.js: each subagent live-tail watcher used fs.watchFile, which stat-polls the file once per second per watcher, forever. Teardown via fs.unwatchFile(path) was fragile and watchers could accumulate across a long app session. Switch to event-driven fs.watch + 300ms debounce, and store a per-watcher teardown() closure called from both stop-subagent-watch and the window-closed handler so nothing leaks. Falls back to a 10s poll only if fs.watch fails to attach. - public/app.js: pollActiveSessions ran every 3s unconditionally (IPC + full-sidebar querySelectorAll) even with zero running sessions. Poll adaptively: 3s while sessions run, 30s when idle. In-renderer session starts re-arm the fast cadence immediately; the 30s floor still catches externally-started sessions. The 30s timeago interval now no-ops when nothing is active. - schedule-runner.js: scanSchedules re-read 4KB of every project JSONL every 60s just to extract projectPath. Resolve it from the cache_meta SQLite cache (getAllFolderMeta) instead, falling back to the JSONL read only when a folder is genuinely uncached. Lint clean; existing 119-test suite passes unchanged.
…-polling perf: cut idle CPU from leaked watchers and unconditional polling
--worktree creates a fresh isolated git worktree, which only makes sense when STARTING a session. On resume (isNew === false) it makes claude try to spin up a new worktree and fail to attach — so a plain sidebar click, and equally the "Create scheduled task" flow (launchScheduleCreator also resumes via openTerminal with isNew=false), silently broke. Rather than stripping the option at each renderer call site, gate it in the open-terminal handler: only append --worktree when isNew. That closes every current and future resume path in one place, and avoids mutating the caller's options object.
…g missing The session cache is only rebuilt from the filesystem on a cold start (populateCacheViaWorker runs only when the cache is completely empty). Once it has any rows, a project folder that changed while the app was closed — or that predates the build which first indexed it — is never re-scanned, so its sessions (and whole worktrees) silently disappear from the sidebar. The live fs.watch only covers changes while the app is running. When the cache is already populated, get-projects now reconciles first: the dead populateCacheFromFilesystem() (defined + exported, never called) becomes a stat-gated reconcileCacheFromFilesystem() that re-indexes only folders that are new or whose newest transcript is newer than cache_meta.indexMtimeMs — cheap when nothing changed. Also extend getFolderIndexMtimeMs to stat the full session-file set (enumerateSessionFiles), not just top-level *.jsonl, so a folder whose only offline change was a subagent transcript under <id>/subagents/ is still flagged for reconcile. Adds regression tests for the gate and the subagent-only case. Rebased onto current main, which wires touchCachedModified/replaceSessionMetrics into the ctx.db literal.
The 2026-06-04 incident: a chain's final step wrote its text + discrete Enter into a session's PTY but the Enter never submitted — the text sat in the composer. The watcher had no post-write verification: the instant-reply heuristic masked the failure, and the final chain step plus the single-command path had no observation at all. Add submitWithVerify(): submit, poll for the busy rising edge within SWITCHBOARD_SUBMIT_VERIFY_MS (default BUSY_RISE_TIMEOUT_MS), and if no rise arrives write a single bare '\r' retry (a no-op on an empty composer, so harmless if the first submit worked) before polling again. The combined waitForTurnComplete is split into submitWithVerify (Phase 1, busy-rise + retry) and waitForBusyFall (Phase 2). Wired into every chain step (the observed rise replaces the old Phase-1 wait — no double wait), the final chain step (verify only, no busy-fall), and the single-command path. The per-step timeout_ms still bounds the whole step, preserving the existing mid-chain timeout semantics. Instant-reply behaviour is preserved: when no rise is confirmed the watcher still proceeds; submit_retries traces it. Result JSON gains additive submit_retries (top-level for single, per-step for chain). Existing assertions updated for the retry '\r' (silent-ctx tests count submissions by command text, CHAIN-12 timing) and 4 new tests cover the verify/retry/no-retry paths.
…bmit-verify fix(trigger-watcher): verify submission and retry the Enter once
…erd) Corrects the 2026-05-31 claim. Witnessed 2026-06-04: cp at 17:17 triggered appimagelauncherd desktop re-integration at 17:17:33; the running instance was cleanly terminated at 17:17:48 (systemd scope end, no segfault, no kernel trace; user confirmed they did not quit). The same swap had survived twice on 2026-06-02 — non-deterministic. Also promotes --config.npmRebuild=false from investigation candidate to field-proven safe-build path (3 successful builds with the instance live).
…er-not-safe docs: cp over a running AppImage is NOT reliably safe (appimagelauncherd)
Don't apply the worktree default when resuming a session
Reconcile cache with filesystem on get-projects so sessions stop going missing
…ll site Review follow-ups: loadProjects() fires get-projects twice per sidebar paint (showArchived false/true), running the reconcile sweep back-to-back; a 1s throttle skips the redundant second pass while the live watcher covers anything landing inside the window. Also note at the call site that the reconcile is synchronous, so the missing await next to the cold-start branch's 'await populateCacheViaWorker()' is intentional.
polish(reconcile): throttle back-to-back sweeps, document the sync call site
…ine expiry When a chain trigger's global deadline fires with unsent steps remaining (starvation via parallel cron re-busying the session), write a new trigger file for the remaining steps with wait:idle so they execute when the session next becomes available. - MAX_REQUEUE = 2 cap prevents infinite re-queue loops - requeue_count field on triggers (validated: non-negative integer) - All four chain-timeout sites re-queue: initial-idle-wait, before-step deadline, step-verify timedOut, step busy-fall timedOut - Timed-out result gains requeued:true + requeue_trigger filename (or requeue_exhausted:true at cap); ok:false/partial:true shape preserved - 8 new tests covering: timeout-with-remaining, initial-wait requeue, cap exhaustion, validation (negative/float/string/absent), success no-requeue Fixes witnessed starvation: trigger e252d69c step 1 never sent, result chain timeout with steps_completed:0.
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
2 participants
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.
Summary
wait: idleso they are processed when the session next becomes freeMAX_REQUEUE = 2cap prevents unbounded retry loops;requeue_countfield tracks depth and is validated (non-negative integer, absent → 0)ok: false, error: 'chain timeout', partial: true) withrequeued: true + requeue_trigger: <filename>extras, orrequeue_exhausted: trueat capTest plan
Fixes witnessed starvation: trigger e252d69c — step 0 (/compact) sent, step 1 never sent due to parallel */4 cron re-busying the session, result {ok:false, error:"chain timeout", partial:true, steps_completed:0}.