Resolves #17329

Mission control gains its WHAT axis: a resident Tasks tab directly beside Activity renders what the deployment is doing — Running, Queued · next, Recent — as provenance-labeled rows reduced server-side from the existing truth verbs. One read-observe wire verb, fleetTasks, reads the orchestrator's deployment-state snapshot (tenant repo sync, per-repo next-due with a backlog gauge, maintenance retry, recovery runs, self-heal freezes), the Memory Core REM backlog, and — where the fleet process can reach it — the Knowledge Base's own ingestion progress; every axis answers as itself, tenant identifiers never leave the reducer, and the ~100 KB snapshot never crosses the wire. The pane projects the envelope through a FleetTask Store, renders a determinate run as a native <progress> element PLUS its number (the bar is the glance, the text is the 1.4.1 channel), keeps a backlog labeled as a queue, carries "in progress" as a first-class word, and keeps the labeled sample on the cold spine. The cockpit drives the read at boot, on every liveness tick, and on reconnect.

Evidence: L3 achieved (real-click whitebox journey through the REAL authenticated transport — loopback bridge double, bearer injected via the product injector, tab click, native progress value/max asserted in a real DOM, refresh intent observed as a second wire request with row replacement; the source reducers are proven against captures of the three live verbs taken on this seat 2026-08-22) → L3 required (AC-2/AC-3 name rendered, click-observable behavior). Residual: none.

AC Evidence

| AC-1 | Outside-CI: the §04 sketch (https://github.com/neomjs/neo/issues/17329#issuecomment-5374041569, 2026-08-21) was accepted by the operator in session 2026-08-22 ("plan sounds solid") before the first line; implementation deltas are appended to that same comment (single artifact). | | AC-2 | CI: test/playwright/unit/apps/agentos/view/fleet/tasksPane.spec.mjs (cold sample, transport-cold sample, source-unavailable empty lines, wired projection per section with progress/backlog/state word/source pill, partial meta line, replacement never accumulates) + test/playwright/unit/ai/services/fleet/fleetTasksSource.spec.mjs (every section from the live-shaped snapshot; the REM backlog gauge from producer-shaped cycles — never a running claim; ingestion active/stalled/idle; ordering + cap). Outside-CI: FleetTasksPaneNL.spec.mjs real DOM (Test Evidence). | | AC-3 | CI: fleetTasksSource.spec.mjs against the PRODUCER contracts per the #17329 Contract Ledger — the reader's real stale and degraded envelopes (ok:false, snapshot retained) keep their rows under their word, the real unavailable envelope and an unknown-status ok:false fail closed, failed/skipped completions are stamped by the current attempt, REM axisErrors on a count axis is the typed rem-axis-error (a fallback zero never becomes a wired measurement) while a recentCycles axis error does not veto the gauge, the unwired ingestion axis never degrades the fold, a throwing axis carries a redacted detail and no rows, viewer binding refused when not canonical — plus fleetCockpit.spec.mjs "the tasks read (loadTasks)" and dispatchFleetRequest.spec.mjs / fleetServer.spec.mjs (the verb on the allowlist and in BOTH policy ledgers as read-observe). Outside-CI: the e2e drives fleetTasks through the authenticated loopback transport; devFleetServer wires the plane operations (get_deployment_state_snapshot, get_rem_pipeline_state) and the in-process-only KB operation. | | AC-4 | CI: fleetCockpitPopOut.spec.mjs (SOUTH_HOME now carries tasks — the generic south-pane pop-out/return witness), fleetCockpitTearOut.spec.mjs (strip order with tasks survives tear-out + return), cockpitDockDocument.spec.mjs (item + stream-tabs order); the pane resolves through getTasksPane() (docked, vesseled, returning) exactly like the memories/wake-routes accessors. | | AC-5 | CI: check-theme-surfaces.yml (path-triggered by resources/scss/src/apps/agentos/**: parity + token-only + completeness). Outside-CI: npm run check-theme-surfaces ✓ and npm run ai:lint-fleet-vocabulary-parity ✓ at this head; TasksPane.scss consumes only --fm-* roles/rhythm/ink tokens and the motion vocabulary — the only literal lengths are the S1 freshness pill's own 2px 7px / 999px (copied verbatim from the AgentDetail/MailboxPane vocabulary so the cockpit keeps ONE freshness language) and 1px hairlines; the progress fill width is data (value/max), never a script-written style. |

Deltas from ticket

1. Queued · next is LIVE, not sample. The sketch (§4) listed "Schedule (next runs): no verb yet — D#11857". Intake falsified that: the deployment snapshot already carries schedule truth — tenantRepoSync.repos[].nextDueAt / due, maintenance.retry.nextAttemptAtMs, and the sweep's task.running / lastCompletion — so queued rows bind to it. The honest sample survives only on the cold spine (no snapshot, or a transport-level fallback with no source axes), never beside live rows.
2. The Knowledge Base ingestion verb is bound in-process mode only. The plane client proves identity against the Memory Core plane (list_permissions) and holds no Knowledge Base session, so plane mode leaves getIngestionProgress out and the source answers that axis as the typed unwired state — which does not degrade the envelope; the snapshot carries the deployment's real ingestion lane in both modes.
3. Row grammar renders the state word AND the optional progress, not one-or-the-other as sketched: the word is the always-present text channel, the bar is additive.
4. The cockpit drives the read at boot, on every liveness tick (with its own in-flight cap) and on reconnect, instead of intent-only: a WHAT surface that says "not observed yet — click Refresh" is not mission control. The pane still never reads on its own (sibling contract).
5. Transport-cold vs source-cold are told apart structurally (no source axes ⇒ the cockpit never reached the source): transport-cold keeps the labeled sample (the Activity stream's fail-closed-never-blank convention); source-cold renders the honest empty lines under the reason.
6. .claude/launch.json gains a cockpit-live preview configuration (npm run cockpit:live) so a plane-bound headed receipt is one click for the next peer.
7. Review round 1 (Euclid) — the reducer now obeys its producers, and the ticket carries the T3 Contract Ledger it lacked. The REM "fresh cycle ⇒ running" branch is gone: HealthService.buildRemPipelineState projects completed-cycle summaries (runId, wallClockMs — a duration, overflow facts, outcome) with no instant and no active flag, so the REM row is always the queue fact under the word backlog, and axisErrors on a count axis answers the typed rem-axis-error instead of a wired zero. The deployment reader's stale / degraded envelopes (ok:false with the snapshot RETAINED) keep their rows under their own word; the real unavailable envelope fails closed with the producer reason. A failed tenant-sync completion is stamped by lastErrorAt, a skipped sweep by lastRunAt — never by the preserved older lastSuccessAt. The first-cut fixtures had modeled fields the producers do not emit; every REM/deployment fixture is now producer-shaped and the ledger names each source of authority.

Test Evidence

• NEO_E2E_PORT=8117 npx playwright test agentos/FleetTasksPaneNL -c test/playwright/playwright.config.e2e.mjs --workers=1 → 1/1 at this head: three live sections, progress[value=100][max=400] + 25%, the backlog gauge 1040 / 2000 under the word backlog with the honest — time, source pills, the meta line naming every axis, and Refresh → second fleetTasks wire request → rows replaced.
• Headed cold-spine receipt on this seat's dev cockpit (npm run cockpit, in-process boot, no bearer landed): the Tasks tab sits second in the south strip and renders the sample-labeled shape under the named transport reason (Tasks unavailable · fleet tasks read failed — the rows below show the shape, not the deployment.), styled through the rebuilt dev theme; computed-style probe: root transparent, rows --fm-panel on --fm-line-soft, progress track --fm-panel-2 at 4px, Refresh reachable.
• npm run check-theme-surfaces ✓ · npm run ai:lint-fleet-vocabulary-parity ✓ · npm run generate-docs-json for the three new classes.
• Mutation diagonal: dropping the repo.disabled === true skip in extractDeploymentRows turns the source witness red (2 failed / 22 passed at the first cut), restored — the guard fails on the defect it names.
• Reviewer falsifiers from round 1, now encoded as red controls in the source spec: axisErrors.undigested with the producer's numeric fallback (was wired, now rem-axis-error); producer-shaped cycles can never flip to "in progress"; the reader's ok:false stale envelope keeps its 7 rows; a fresh failed completion stamps the failed run, not the older success.

Post-Merge Validation

None owed: the contract (reduction, envelope fold, projection, wire, parity) is proven in-tree and through the real transport. A plane-bound headed cockpit adds environment coverage only; this seat's dev cockpit boots every pane without a bearer today (handshake endpoint healthy, redemption not landing — defect-note sent), which is the launcher's lane, not this pane's.

Commits

• b47ef7119d the verb, source + wiring, model + store, the pane + skin, cockpit loader/factory/accessor/tick, dock document, policy ledgers, both wire lists, docs json, six spec updates + two new unit witnesses (rebased onto dev 4cfeb3858d; the only conflict was the generated docs/output/class-hierarchy.json, regenerated).
• f10215e16e the whitebox journey FleetTasksPaneNL.spec.mjs through the authenticated loopback transport.
• 93e1734291 review round 1: the reducer obeys its producers (REM axisErrors + instant-less cycles, the reader's stale/degraded/unavailable envelopes, completion times by status), producer-shaped fixtures + the reviewer's red controls, the degraded word on the pane.
• 8b02bc42f8 the returns record drops the TS-optional syntax the docs parser rejects (check-jsdoc-types clean; docs json regeneration produced no diff).

Authored by Clio (Claude Fable 5, Claude Code). Session 14acab5a-4b6c-4987-91c7-f683e39baa55.

Author response — round 1

Receipts: fleetTasksSource spec 28/28 · owning unit trees unit/ai/services/fleet unit/apps/agentos 1496 passed · no change to the wire lists, policy ledgers, pane projection or the e2e journey (its fixture envelope already carried the REM row as an instant-less queued gauge). Exact-head CI running; re-seat on green.

— Clio (Claude Fable 5, Claude Code). Session 14acab5a-4b6c-4987-91c7-f683e39baa55. 📜

@neo-fable-clio commented on 2026-08-22T14:59:38Z

Author response — round 2 (RA-3 residual)

Your drift finding was exact: the ledger's envelope row said wired/stale while the code's ANSWERED_STATES says wired|stale|degraded, and no envelope-level witness resolved it. Three touches at b9b6c7f290: the #17329 ledger's fold row now names wired, stale or degraded as answered (matching ANSWERED_STATES by name); a new readTasks-level witness proves a degraded deployment envelope folds to wired with the axis carrying its own word and its rows surviving; source spec 29/29. Matrix, code and witness agree exactly — no other file touched. CI running at exact head; re-seat on green.

— Clio (Claude Fable 5, Claude Code). Session 14acab5a-4b6c-4987-91c7-f683e39baa55. 📜