pi-crew 0.10.6 → 0.11.1

package/CHANGELOG.md

@@ -2,6 +2,221 @@
 
 > **Note:** `atomic-write-v2.ts` / `AtomicWriter` mentioned in historical entries below was consolidated into `atomic-write.ts` as of v0.9.42. This changelog is preserved as historical record — the migration was completed (the v2 class was never adopted; v1 won on simplicity + symlink-safety + link+unlink atomicity). See `docs/migration/atomic-write-v2-migration.md` for the decision rationale.
 
+## [0.11.1] — RAIL: one UI design language for every surface + UI-audit fixes (2026-09-16)
+
+### feat(ui): RAIL — a single visual language for the whole product
+
+- **New shared design system.** `src/ui/rail.ts` owns every glyph and helper —
+  rail `┏ ┣ ┃ ┗`, the `NAME ▸ SUBJECT` canopy, dot leaders `······`, the
+  eighth-block gauge `▕████▎░░▏`, status slot/badge/icon maps, `overflowHint`,
+  `formatHint`/`keyToken`, `CURSOR`/`ACTIVE` — and `docs/UI-DESIGN-SYSTEM.md`
+  documents the grammar, the seven surface classes and the width contract.
+  `src/ui/tool-renderers/index.ts` now imports from it instead of declaring its
+  own copies.
+- **Tool card (team + subagent) redesigned**: width-deferred through
+  `AdaptiveCard` (a frame built for 116 columns no longer tears at 100), one
+  unified identity canopy (`┏ CREW ▸ <team>`, `┏ AGENT ▸ <name>`,
+  `┏ LIST`), agent name surfaced while the run is still streaming, fail-visible
+  outcome on three levels, `done`/`succeeded` alias handling, merged
+  tally+gauge with the elapsed segment right-aligned, and `role/agent`
+  duplication collapsed (`verifier/verifier` → `verifier`).
+- **Every other surface migrated**: crew widget (still exactly ONE row:
+  `┃ ⠧ CREW ▸ fast-fix · 2 running · 3/5 done · ⏰ 1 sched ···· ↓·enter`, idle row
+  `┃ CREW ▸ idle · …`, focused `❯` marker), plan/task list (`┏ PLAN ▸ <title>`),
+  live-run sidebar (`┏ LIVE ▸ <runId8>` + `┣ ACTIVE ▸ n`), team dashboard,
+  Agents & Jobs browser, settings overlay, transcript viewer, live-conversation
+  / help / confirm / mailbox-detail / mailbox-compose / agent-picker overlays,
+  dashboard panes and `/team-mascot`. The rounded box `╭─╮│╰─╯├┤` and the inline
+  `── label ──` rule are retired (sections are `┣ NAME ▸ x`).
+- **One dialect per concept**: cursor `›`, active/section marker `▸`, overflow
+  `▲ n above` / `▼ m below` (was four dialects), hints `keys label · …` with the
+  close/cancel action last and canonical key spelling (`Esc`, `Enter`, `Tab`) —
+  including an `exactKeys` mode for the case-sensitive plan-approval keyspace
+  (`A` approve vs `n` deny).
+
+### fix(ui): bugs found by the audit, the cold-verify pass and live runs
+
+- **`undefined — ↓·enter` in the widget after a run finished** — the zero-runs
+  branch built its row without checking `schedLine`; with no active runs and no
+  schedules it painted the literal `undefined`. It now renders nothing.
+- **No `undefined` can reach a string**: every interpolated field on records read
+  from disk (`agents.json`, schedules, metrics) is guarded — role/agent/model
+  routing in the sidebar and agent picker, every persisted job field in the
+  schedules pane, metric label values, and the settings `Edit` title.
+- `1 tools` → `1 tool` (one `formatCount` helper, five surfaces).
+- Help cheatsheet no longer prints a raw TAB (`keyToken` missed `\t`).
+- `role->agent` → `role▸agent` everywhere, including the producer
+  (`task-display.ts`) rather than only at one render boundary.
+- Settings overlay: the edit submenu hint is built by `formatHint` and no longer
+  advertises a key that is not bound.
+
+### fix(ui): phantom affordances, schema drift and theme discovery (P0)
+
+- **`/team-settings theme <name>` works again**: `theme-discovery.ts` used bare
+  `require("node:fs")` inside a module loaded by dynamic import, so discovery
+  could fail silently and only the builtin themes were selectable. Now a
+  top-level import, with a HOME-sandboxed test (13 themes discovered, active
+  theme read from `settings.json`, `setPiTheme()` writes).
+- Removed the phantom `p` (progressToggle) binding/action/help entry, the no-op
+  `placement` option, and the non-existent `/team-cleanup` / `/team-health`
+  commands from help.
+- Schema/docs drift: `ui.widgetRowStyle`, `ui.inlinePanel`,
+  `ui.autoCloseDashboardMs` added to `schema.json` + configuration schema so
+  editors stop flagging valid keys; brief-mode command description scoped to
+  what it really does; help overlay gained the Schedules and Plan key groups;
+  keyboard reference and command docs corrected.
+- `LiveConversationOverlay` gained a real `handleInput` (scroll, auto-scroll
+  toggle, close) so its advertised footer keys are reachable.
+- **Terminal status restored**: the documented tab-title + Ghostty `OSC 9;4`
+  progress integration is now wired (`installTerminalStatus`) instead of the
+  no-op dispose handler; best-effort, debounced, headless-safe.
+
+### chore: dead code removed
+
+- Deleted `overlay-stack.ts`, `ui/loaders.ts`, `crew-vibes/figures.ts`,
+  `crew-vibes/font-detect.ts`, `dynamic-border.ts` and `boxLine()` plus their
+  tests (~1.7k lines net), together with the dead `compactAgentRow` chain.
+- Removed the orphaned `ui.widgetRowStyle` config chain (default, type,
+  parser, TypeBox schema, `schema.json`, KNOWN_KEYS, `model.rowStyle` field,
+  `WidgetRowStyle` type): the M4 RAIL dock renders exactly one layout, so the
+  key had no reader left. The strict schema now rejects it as unknown;
+  schema-sync + settings tests pin the removal in both directions so it cannot
+  resurface.
+
+### fix(ui): defects found by rendering a REAL team run
+
+Rendering the live run `team_20260916110100` through every surface (tool card,
+widget dock, plan card, sidebar, dashboard, agents & jobs browser) exposed seven
+defects the review passes had missed:
+
+- **The dock spun forever.** `┃ ⠹ CREW ▸ fast-fix · 0 running · 3/3 done` — a
+  spinner with nothing to spin for, because the header always received a braille
+  frame. The leading glyph now reflects state (`⠋…` only while a worker is
+  running, otherwise `✓`/`✗`); zero counts (`0 running`) are no longer printed.
+- **The dock lost its key hint on a narrow terminal.** At 50 columns the row
+  ended in `···· ↓…`; the leader budget now shrinks with the width so the left
+  segment is trimmed (with `…`) and `↓·enter` always survives.
+- **The run dashboard truncated the status into `complete…`.** The narrow-width
+  fallback clipped `head · meta` as one string, cutting the very token the run
+  list exists to show. Meta now yields first, then the goal; the head (run id +
+  status) is only clipped if it cannot fit at all.
+- **Sidebar printed an invented `· right default` string** — replaced with the
+  run's real `workspaceMode` (`single` / `worktree`).
+- **`fast-fix/fast-fix`** in the sidebar and dashboard — one shared
+  `teamWorkflowLabel()` collapses the workflow half when it repeats the team.
+- **Raw usage rows** (`input=2780, output=3715, cacheRead=57216, cost=0.000000,
+  turns=0`) — the sidebar now uses the compact `↑2.8k ↓3.7k` form the tool card
+  already had (the `key=value` form stays for CLI/status output).
+- **`314.7s` instead of `5m44s`**, `1 runs` instead of `1 run`, `3 agents` /
+  `N runs` without pluralisation, and `┃ - none` for empty sections — all
+  normalised through `formatDuration` / `formatCount`.
+
+### verification
+
+- `test:unit` full suite, `test:critical` (117/117), `test:bundle` (2/2), biome,
+  `format:check`, `tsc --noEmit` and bundle size/staleness gates all run on the
+  release tree; UI migration verified by an independent verifier rendering real
+  surfaces at 40/60/80/100/160 columns.
+
+## [0.11.0] — Scheduled Jobs UI + Agents & Jobs browser + kitty CSI-u hardening (2026-09-15)
+
+### feat(ui): Agents & Jobs browser — the crew widget's ↓+Enter popup
+
+- The crew widget is now ONE screen line (`Crew agents 3/5 · 1 running · 24m · ↓·enter`); pressing `↓` or Enter opens the Agents & Jobs browser — a framed modal (border box with live counts in the top frame, key hints in the bottom frame) listing every widget-visible agent and scheduled job from the SAME data source as the widget counts (`activeWidgetRuns()` over `.crew/state/runs/*/agents.json` + `recordTokPerSec()` usage snapshots), so the browser and the widget can never disagree.
+- Agent rows show role/status/duration/tok-per-sec live; job rows show cron/interval + next run. Focus modes cycle Agents → Jobs; jobs keep their inline detail view.
+- Pressing Enter on an agent row opens the full transcript overlay (`DurableTranscriptViewer` — the same experience as `/crew transcript`). For RUNNING agents the transcript path resolves through a 3-tier fallback (record.transcriptPath → status.json → artifacts-root `transcripts/<taskId>.attempt-N.jsonl` convention), because records only carry the path after completion.
+- The legacy `p` key (spawn a mux pane tailing the watcher-formatted log) is retired: a log tail is not a real pi session. The retired code stays as commented blocks for a future "focus the existing surface pane" key.
+
+### feat(agents): live surface panes — watch a worker's REAL pi TUI
+
+- New `runtime.surface` config (`~/.pi/agent/pi-crew.json`): `{ mode: "herdr" | "tmux", visibleAgents: ["agent", ...] }` boots eligible workers inside a real mux pane running the pi TUI (the worker's `--mode json -p` args are stripped for the pane) — watch the worker's actual footer/model/widget live instead of log tails. The pane closes itself when the worker finishes; resume is never used (resuming would kill the subagent).
+- Surface spawns emit `worker.surface_spawned` (+ `surfacePaneId` on `worker.started`); gate rejections emit `worker.surface_gate_blocked` with the reason. Async workers forward TERM/HERDR_*/mux env into the child.
+
+### fix(ui): kitty CSI-u hardening — matchesKey everywhere
+
+- Every overlay now routes Enter/Escape through `@earendil-works/pi-tui`'s `matchesKey` instead of raw byte compare: kitty's disambiguate-codes protocol sends `\x1b[13u`/`\x1b[27u`, so raw `\r`/`\x1b` compares silently swallowed keys. Sweep covered the agents browser, confirm, agent-picker, mailbox detail, mailbox compose, live-conversation overlays, transcript viewer, mascot, and settings overlay (backspace ×2).
+- `src/ui/key-utils.ts` re-exports `matchesKey` for overlays that cannot import pi-tui directly.
+
+### fix(ci-windows): the two real flakes behind windows-latest failures
+
+- Broker integration helpers compressed EVERY client timer to 100ms (intent: only the 1s hello deadline) — the 15s request timeout and the explicit 2s escalate budget fired before slow NTFS durable writes could land, nondeterministically killing clients with fallback-sticky `request-timeout` (broadcast/DM/escalate tests). Timers ≤1s compress; real request budgets stay real.
+- `O_CREAT|O_EXCL` against a `.flock` held mid-create by another process surfaces as EPERM/EBUSY (not EEXIST) on Windows — both lock retry loops now treat EPERM/EBUSY as contention (deadline still bounds the wait). Fixes the ST-3 cross-process flock EPERM crash.
+- Also in this release: `npm run lint` (biome) added to the pre-push gate — three unused members left by the p retirement had slipped past a tsc+format-only gate.
+
+### Scheduled Jobs UI (tiers A/C/D/E — previously unreleased)
+
+### feat(vibes): tok/s speed UI REMOVED — pi's built-in working indicator always used
+
+- Maintainer decision: the entire speed subsystem is deleted, not toggled off — custom spinner frames (`setWorkingIndicator`), the `Working N tok/s` message (`setWorkingMessage`), SpeedTracker/SpeedAnimator, the speed config block, and every speed event handler (agent_start/turn_start/message_*/turn_end/agent_end wiring). crew-vibes is now a single-purpose module: fetch the provider rate-limit quota and publish it as a status entry on pi's native footer. `/team-vibes` shrinks to `[on|off]`.
+
+
+### feat(footer): restore pi's NATIVE footer — custom footer replacement retired (UI-review option 1)
+
+- Maintainer decision: the crew-vibes custom footer (setFooter replacement that re-implemented pi's pwd/stats/status lines to guarantee the quota a never-truncated line) is REMOVED — pi's built-in footer is always shown. The re-implementation had drifted from native (missing `CH%` cache-hit, missing `xp` segment, totals ignoring toolResult/branch-summary/compaction usage, hardcoded `(auto)`), and composing pi's native FooterComponent is impossible by design (extensions only receive ReadonlySessionManager, not AgentSession).
+- Provider quota (`z.ai 5h ━ 37% 46m Wk …`) now publishes through `ctx.ui.setStatus` and joins pi's native status line (right-truncates on very narrow terminals — accepted).
+- The crew widget ALWAYS renders through pi's widget slots now (`bottom` placement maps to belowEditor): FooterDockHost, the dock-footer registry, slotInstalled/dockedInFooter state, and WidgetRenderOptions.noSchedulesLine are deleted. The ⏰ schedules line and all run rows paint in the belowEditor slot exactly as the pre-footer design; the keep-alive fix is preserved.
+- Clock-injection sweep (D6-T4) in the render path: compactDockLines/detailed rows/agents-pane now pin ONE clock per render (options.now / options.nowMs) threaded into orderWidgetAgents, dockUsageText, agentStats, computeLiveDurationMs — no more scattered Date.now() reads that could straddle a tick.
+
+
+### feat(footer): retire the capacity stage meter; the ⏰ schedules segment takes its slot
+
+- Maintainer decision after the full UI review: the capacity meter (context token count `74k` + `○ Orbit/Cruise/Warp/…` stage glyph) duplicated the context percent already shown on the stats line and carried no actionable signal — removed from the footer composition (`renderCapacity` remains exported for its unit tests).
+- The meter line now carries the Tier-C schedules segment on the LEFT (`⏰ N sched · next …`) with the provider quota right-aligned on the SAME line (wraps to two lines only on narrow terminals). Net effect: one screen line saved — the schedules line no longer paints as its own dock line when the crew-vibes footer is active.
+- `CrewWidgetModel.dockedInFooter` + `WidgetRenderOptions.noSchedulesLine`: in the dock path the widget no longer paints `⏰` itself (the footer owns it — no duplicate painter); slot mode (no footer sink) keeps painting it unchanged.
+
+### fix(scheduler): cron jobs never fired — arm() had no `cron` branch (P1-1 from UI review)
+
+- Root cause of the live "next 7h ago" symptom: `arm()` only handled `interval`/`once`, so cron jobs were persisted with a `nextRun` but never got a timer — silent death. `nextRunTime()`/`nextCronDate` (already in the same file) were simply never wired into arming.
+- `arm()` now arms cron via a chained `setTimeout` hop toward the next occurrence, clamped below the 2^31-1 ms timer ceiling (`MAX_TIMER_DELAY_MS`) so yearly crons cannot mis-fire on overflow; `disarm()` clears the cron handle on the right branch.
+- After each cron fire, `advanceCronNextRun()` recomputes and persists the next occurrence (previously `nextRun` was written exactly once, at job creation); jobs whose next occurrence is not computable self-disable with an `error` event explaining why.
+- Scheduler constructor takes an injectable clock (`now?: () => Date`, default `Date.now`) per the D6-T4 clock-injection convention — cron firing tests run on a fake clock, no real sleeps.
+- New suite `test/unit/runtime/scheduling/scheduler-cron-arm.test.ts` (first fire, re-arm + nextRun advance, mid-flight disable, overflow chaining, disarm) — red on the pre-fix HEAD, green with the fix.
+
+### feat(ui): B2 gate hint — "N project-tier jobs hidden" on all three surfaces (P2-1)
+
+- The Wave B2 opt-in gate (`schedulingEnabled` + `allowProjectScheduledJobs`, both required in user-tier `~/.pi/crew-settings.json`) was completely invisible: gated-out jobs made every surface fall into the plain empty state, looking like the jobs were lost.
+- New companion getter `getScheduledJobsHiddenCount()` / pure `scheduledJobsHiddenCountOf(tiers)` in `settings-store.ts` (no signature change to `getScheduledJobs`); the paint path reuses the registration-time tiers stash — no new disk reads on render.
+- Schedules pane + `/schedules` (via the shared renderer, parity test kept) append one dim hint line, the widget line appends a compact `· N hidden` — all three surfaces, EN-only per project convention, shown only when hiddenCount > 0.
+
+### test: settings-store tests are now hermetic
+
+- `loadCrewSettings()` gained an optional `globalFile` param (default unchanged). The three affected tests previously read the developer's REAL `~/.pi/crew-settings.json` ("defaults when file missing" failed the moment that file exists with scheduling flags) — they now pin an absent temp global file.
+
+Full Scheduled Jobs UI stack over the existing scheduler/settings layer. No version bump yet — no schema or persistence changes; reads a single provider, mutations go through the existing extension channel.
+
+### feat(ui): schedules dashboard pane (tier A)
+
+- New pane 8 in the run dashboard (`src/ui/dashboard-panes/schedules-pane.ts`, pure string renderer): one main line per job (enabled glyph ●/○ · name truncate · humanized schedule · relative next-run · last status ✓/✗/⟳ · runCount) plus a sub line (subagentType · lastRun · id in headless mode). Empty state: `No scheduled jobs — create via team tool action='schedule'`.
+- Keys `T/N/V/X/R` (pane-scoped in `keybinding-map.ts`): toggle-enabled, run-now, details toggle (goal + ScheduleSpec + spawnedRunIds), delete with a **2-step confirm-gate** (first X arms + warning line, any other key disarms), refresh (drops the provider cache). Job cursor up/down while pane 8 owns input.
+- Mutations never touch the scheduler from the UI layer: they leave the dashboard only as `done()` selections carrying `schedule-*` actions + `jobId`, routed by `commands/shared.ts openTeamDashboard` through `handleTeamTool({action:'schedule', subAction, jobId})` → `handle-schedule.ts` (the same channel as the chat tool), then the result surfaces as an info/error toast and the dashboard reopens.
+
+### feat(ui): schedules line in the crew widget (tier C)
+
+- One lowest-priority widget row (`src/ui/widget/widget-renderer.ts`): `⏰ N sched · next Xm` — painted ONLY when ≥1 enabled job exists, always last (below active-run info). The line joins the widget cache signature so add/remove/toggle repaints immediately. Survives the no-runs collapse (scheduled jobs are exactly what runs while nothing interactive is active). One clock read per render feeds both the signature and the painted line.
+
+### feat(extension): scheduler event toasts (tier D)
+
+- New `schedule-toast-bridge.ts` wired into the scheduler's `emit` in `lifecycle-handlers.ts`: `fired` → `⏰ <name> fired → <agentId>` (info); failures → red. Failure has two producers — the scheduler's synchronous `error` event and the (common) async executor rejection recorded as a `running→error` lastStatus transition — deduped per attempt so one failed attempt = one notice. Success surfaces as `⏰ <name> ✓ succeeded` on a tracked lastStatus transition only (no drip on patches). Bounded like the hung-notice pattern; headless no-ops; all interpolated fields sanitized.
+
+### feat(commands): /schedules + /schedules log (tier E, headless-compatible)
+
+- New `/schedules` command re-uses the SAME pane renderer (text-block variant — no copied table logic) so command and dashboard can never drift. Natural-language phrases `crew schedule` / `scheduled jobs` rewrite via the crew-input-router.
+- `/schedules log <jobId-or-name>` resolves a job by id or name, tails the most recent output artifact of its latest spawned run (32KB bound, traversal-guarded via `resolveRealContainedPath`; unsafe-charset runIds degrade to an error result). Works under `pi -p` — the handler touches only `ctx.cwd` + `ctx.ui.notify`.
+
+### fix(scheduling): run-now on a once job consumes it
+
+- `CrewScheduler.runNow()` on a `once` job now self-disables after the forced fire (mirroring the timer-driven path). Previously the still-armed timer would fire the job a SECOND time at its scheduled time — a one-shot executing twice.
+
+### Constraints honored
+
+- **Single source of truth (G17)** — every consumer (pane, command, widget, autocomplete) reads jobs through one provider `getScheduledJobs()` (scheduler singleton, gated settings-tier fallback). No defaults copies anywhere.
+- **Injectable clock (D6-T4)** — renderers take `now: Date` as a parameter; no `Date.now()` on any render path (pane, widget, command, and the provider TTL timestamp).
+- **Extension channel only** — toggle/run-now/delete dispatch through `handle-schedule.ts` subActions; persisted `scheduledJobs` shape unchanged.
+
+### Tests
+
+- New: `schedules-pane`, `schedules-dashboard`, `widget-schedules-line`, `relative-time`, `schedules-command`, `schedule-toast-bridge`, `team-tool-schedule-provider`, `scheduler-run-now`, `dashboard-schedule-routing` (155+ assertions across render parity, injected clock, confirm-gate, toast bounds/dedup, traversal guards, once-job run-now regression, shared.ts routing integration).
+
 ## [0.10.6] — agent/skill resource layer + broker coordination fixes (2026-09-12)
 
 ### fix(bundle): PACKAGE_SKILLS_DIR resolves via `packageRoot()` instead of broken `import.meta.url` walk-up

package/README.md

@@ -699,7 +699,6 @@ the `goal` action's completion-guarantee loop to builtin workflows.
 | Key path | Type | Default | Description |
 |----------|------|---------|-------------|
 | `ui.widgetPlacement` | `aboveEditor \| belowEditor \| bottom` | `bottom` | Where the status widget renders. `bottom` docks inside the crew-vibes footer (very bottom of the screen), falling back to `belowEditor` when no footer sink exists. |
-| `ui.widgetRowStyle` | `compact \| detailed` | `compact` | Per-agent row layout in the widget (`compact` = one width-budgeted line per agent; `detailed` = two-line tree). |
 | `ui.inlinePanel` | `boolean` | `true` | Keyboard-navigable agent rows under the prompt (`↓` from an empty prompt). Yields to any extension that owns the editor component. |
 | `ui.refreshMs` | `number` | `1000` | Internal UI refresh cadence default (not a user-settable `ui.*` schema key — kept here for completeness). |
 | `ui.widgetDefaultFrameMs` | `number` | `1000` | Internal widget frame-budget default (not a user-settable `ui.*` schema key — kept here for completeness). |