T1750: Taskcore control plane fixes — metadata, queue, decomposition, cleanup

.task

@@ -0,0 +1,10 @@
+{
+  "taskId": "1750",
+  "phase": "execution",
+  "fenceToken": 9,
+  "sessionId": "2c16cdd1-6e26-4b9a-9aab-0f83c2d423ab",
+  "journalPath": "/tmp/taskcore-worktrees/journal-T1750/tasks/T1750/",
+  "codeWorktree": "/tmp/taskcore-worktrees/code-T1750",
+  "claimedAt": 1773508861049,
+  "reviewNotes": []
+}

core/test/validator.test.ts

@@ -458,3 +458,127 @@ test("TaskReparented: reparent of terminal task succeeds", () => {
   const error = validateEvent(state, event);
   assert.equal(error, null);
 });
+
+function reviewActiveState(): SystemState {
+  let state = createInitialState();
+  const events: Event[] = [
+    {
+      type: "TaskCreated",
+      taskId: "RV1",
+      ts: 1,
+      title: "Review validation",
+      description: "Review fence coverage",
+      parentId: null,
+      rootId: "RV1",
+      initialPhase: "execution",
+      initialCondition: "ready",
+      attemptBudgets: { analysis: { max: 2 }, decomposition: { max: 2 }, execution: { max: 2 }, review: { max: 2 } },
+      costBudget: 5,
+      dependencies: [],
+      reviewConfig: { reviewers: ["overseer"], policy: "all" },
+      skipAnalysis: true,
+      metadata: {},
+      source: { type: "middle", id: "test" },
+    },
+    {
+      type: "LeaseGranted",
+      taskId: "RV1",
+      ts: 2,
+      fenceToken: 1,
+      agentId: "coder",
+      phase: "execution",
+      leaseTimeout: 60_000,
+      sessionId: "exec-1",
+      sessionType: "fresh",
+      contextBudget: 512,
+    },
+    {
+      type: "AgentStarted",
+      taskId: "RV1",
+      ts: 3,
+      fenceToken: 1,
+      agentContext: {
+        sessionId: "exec-1",
+        agentId: "coder",
+        memoryRef: null,
+        contextTokens: 128,
+        modelId: "test",
+      },
+    },
+    {
+      type: "PhaseTransition",
+      taskId: "RV1",
+      ts: 4,
+      from: { phase: "execution", condition: "active" },
+      to: { phase: "review", condition: "ready" },
+      reasonCode: "work_complete",
+      reason: "work_complete",
+      fenceToken: 1,
+      agentContext: {
+        sessionId: "exec-1",
+        agentId: "coder",
+        memoryRef: null,
+        contextTokens: 128,
+        modelId: "test",
+      },
+    },
+    {
+      type: "LeaseGranted",
+      taskId: "RV1",
+      ts: 5,
+      fenceToken: 2,
+      agentId: "overseer",
+      phase: "review",
+      leaseTimeout: 60_000,
+      sessionId: "review-1",
+      sessionType: "fresh",
+      contextBudget: 512,
+    },
+    {
+      type: "AgentStarted",
+      taskId: "RV1",
+      ts: 6,
+      fenceToken: 2,
+      agentContext: {
+        sessionId: "review-1",
+        agentId: "overseer",
+        memoryRef: null,
+        contextTokens: 128,
+        modelId: "test",
+      },
+    },
+  ];
+
+  for (const event of events) {
+    const reduced = reduce(state, event);
+    assert.equal(reduced.ok, true, `Event ${event.type} failed: ${!reduced.ok ? reduced.error.message : ""}`);
+    state = reduced.ok ? reduced.value.state : state;
+  }
+
+  return state;
+}
+
+test("ReviewVerdictSubmitted rejects stale review fence token", () => {
+  const state = reviewActiveState();
+  const event: Event = {
+    type: "ReviewVerdictSubmitted",
+    taskId: "RV1",
+    ts: 7,
+    fenceToken: 1,
+    reviewer: "overseer",
+    round: 1,
+    verdict: "approve",
+    reasoning: "Looks good",
+    agentContext: {
+      sessionId: "review-1",
+      agentId: "overseer",
+      memoryRef: null,
+      contextTokens: 128,
+      modelId: "test",
+    },
+  };
+
+  const error = validateEvent(state, event);
+  assert.ok(error);
+  assert.equal(error.code, "stale_fence_token");
+});

docs/daemon-post-endpoints.md

@@ -0,0 +1,50 @@
+# Taskcore Daemon POST Endpoints Reference
+
+Source: `middle/http.ts`
+
+## Endpoint Summary
+
+| Route | Purpose | Key Body Fields |
+|-------|---------|-----------------|
+| `POST /tasks` | Create task | title*, description*, assignee, reviewer, priority, parentId, dependsOn, costBudget, skipAnalysis |
+| `POST /tasks/:id/events` | Raw event | type*, (any core event fields) |
+| `POST /tasks/:id/status` | Status transition | status* (review/done/blocked/pending/execute/decompose/cancel), evidence, blocker, stateRef |
+| `POST /tasks/:id/reparent` | Reparent task | newParentId* |
+| `POST /tasks/:id/revive` | Revive failed/blocked | phase, resetAttempts, reason |
+| `POST /tasks/:id/budget` | Increase budget | attemptBudgetIncrease, costBudgetIncrease, reason |
+| `POST /tasks/:id/decompose/start` | Begin decomposition | (none) |
+| `POST /tasks/:id/decompose/add-child` | Add decomp child | title*, description*, costAllocation*, skipAnalysis, assignee, reviewer, dependsOnSiblings |
+| `POST /tasks/:id/decompose/commit` | Finalize decomp | approach |
+| `POST /tasks/:id/decompose` | One-shot decomp | children* (array), approach |
+
+Also: `PATCH /tasks/:id/metadata` — update metadata fields (priority, assignee, etc.)
+
+\* = required
+
+## Status Transition Map
+
+```
+status="execute":   analysis.active     → execution.ready
+status="review":    execution.active    → review.ready
+status="done":      review.active       → terminal:done
+status="pending":   review.active       → execution.ready  (changes requested)
+status="blocked":   any non-terminal    → terminal:blocked
+status="cancel":    any non-terminal    → terminal:canceled
+status="decompose": analysis.active     → decomposition.ready
+```
+
+## Decomposition Flow (Incremental)
+
+```
+POST /tasks/:id/decompose/start      → creates in-memory session, returns budget
+POST /tasks/:id/decompose/add-child   → adds child spec (repeat)
+POST /tasks/:id/decompose/commit      → creates all children, parent → review.waiting
+```
+
+## Key Behaviors
+
+- **Fencing**: All writes use the task's `currentFenceToken` to prevent stale mutations
+- **Registry validation**: Assignee/reviewer validated against `agents.json`
+- **Cost enforcement**: Decomposition child costs must sum ≤ parent remaining budget
+- **Completion checks**: Tasks with `metadata.repo` require a valid stateRef; parent tasks with `completionRule='and'` require all children done
+- **Notifications**: `done` and `blocked` transitions notify Telegram targets in `metadata.informed`

docs/ops/taskcore/t1750-execution-plan.md

@@ -0,0 +1,122 @@
+# T1750: Taskcore Control Plane Fixes — Execution Plan
+
+**Epic owner:** ceo  
+**Review owner:** kelvin  
+**Status:** Open coordination epic  
+**Updated:** 2026-03-14 heartbeat session
+
+---
+
+## What changed in this update
+
+This plan was corrected to match the **live** task graph rather than the earlier aspirational decomposition draft:
+
+- removed canceled/done children from the active-work list (`T2272` canceled, `T2273` done)
+- restored the **full** metadata-inheritance scope for the active child (`T2271`), based on the earlier `T2267` root-cause analysis
+- documented that **T2268 owns the merged T2268 + T2272 scope** (scheduler/dispatcher + claimability.js/CLI + reviewer deadlock detection)
+- backfilled `parentId`, `reviewer`, and `repo` metadata onto the outstanding children so prompts and worktree provisioning line up with the parent epic
+
+---
+
+## Live child status
+
+| Task | Status | Purpose | Notes |
+|---|---|---|---|
+| T2267 | blocked | Historical metadata-inheritance attempt | Earlier execution was misprovisioned against the wrong repo. Metadata now corrected to point at `/home/ubuntu/taskcore`; use as historical evidence, not as the primary forward work item. |
+| T2268 | analysis.waiting | Queue claimability / review deadlocks | Active child. Includes merged T2272 claimability + CLI scope. |
+| T2269 | analysis.waiting | Decomposition state reconciliation | Active child. |
+| T2270 | analysis.waiting | Attention alerting integration | Active child; depends on T2274. |
+| T2271 | analysis.waiting | Metadata inheritance | Active child. Scope is full parent metadata propagation, not only priority/assignee/reviewer. |
+| T2272 | canceled | Duplicate of T2268 | Do not schedule. |
+| T2273 | done | Completed decomposition slice | Do not schedule. |
+| T2274 | analysis.waiting | Attention formatter crash fix | Active child. |
+| T2275 | analysis.waiting | Obsolete / blocked task cleanup | Active child. |
+
+---
+
+## Active execution waves
+
+### Wave 1 — parallel now
+
+1. **T2271 — Metadata inheritance**
+   - Required propagation set: `repo`, `base_branch`, `informed`, `consulted`, `parentId`, reviewer / assignee / priority defaults, and any other operational metadata that decomposition currently drops.
+   - Why first: fixes future child provisioning / context loading and removes one of the root causes already observed in T2267.
+
+2. **T2274 — Attention formatter crash**
+   - Harden the formatter / endpoint path so null or malformed task data cannot blow up the attention surface.
+   - Why first: prerequisite safety work for T2270.
+
+3. **T2275 — Obsolete / blocked task cleanup tooling**
+   - Build cleanup tooling with dry-run safety and blocker-aware filters.
+   - Why first: operational hygiene without dependency on the other bug fixes.
+
+### Wave 2 — after Wave 1 context is stable
+
+4. **T2268 — Queue claimability false positives + review deadlocks**
+   - Dispatch-side: exclude `review.waiting` / non-claimable states from scheduler availability.
+   - Claim-side: fix CLI/claimability false positives, active-lease visibility bugs, and reviewer-role deadlock detection.
+   - This is the merged successor for T2272's canceled scope.
+
+5. **T2269 — Decomposition state reconciliation**
+   - Make decomposition retries / partial materialization idempotent.
+   - Ensure superseded children are handled safely when a later decomposition version exists.
+
+### Wave 3 — depends on T2274
+
+6. **T2270 — Wire attention output to automated alerting**
+   - Only start once the formatter path is hardened.
+
+---
+
+## Metadata backfill applied during this session
+
+The following outstanding children were patched in live task state:
+
+- `T2267`
+- `T2268`
+- `T2269`
+- `T2270`
+- `T2271`
+- `T2274`
+- `T2275`
+
+Backfilled fields:
+
+- `parentId=1750`
+- `reviewer=kelvin`
+- `repo=/home/ubuntu/taskcore`
+
+Why this matters:
+
+- child prompts can now load the parent journal context
+- review-ready children will not silently deadlock on a missing reviewer
+- future code worktrees will target the real taskcore repo instead of the wrong workspace or no repo at all
+
+Verification snapshot after the backfill:
+
+| Task | `parentId` | `reviewer` | `repo` |
+|---|---|---|---|
+| T2267 | `1750` | `kelvin` | `/home/ubuntu/taskcore` |
+| T2268 | `1750` | `kelvin` | `/home/ubuntu/taskcore` |
+| T2269 | `1750` | `kelvin` | `/home/ubuntu/taskcore` |
+| T2270 | `1750` | `kelvin` | `/home/ubuntu/taskcore` |
+| T2271 | `1750` | `kelvin` | `/home/ubuntu/taskcore` |
+| T2274 | `1750` | `kelvin` | `/home/ubuntu/taskcore` |
+| T2275 | `1750` | `kelvin` | `/home/ubuntu/taskcore` |
+
+---
+
+## Open coordination risks
+
+- **T2267 vs T2271 overlap:** T2267 remains blocked historical evidence while T2271 is the intended forward execution slot. Future submissions must not pretend the blocked historical attempt is the live implementation path.
+- **Immutable task descriptions:** live task descriptions still reflect earlier wording; until taskcore gains a description-edit path, the parent journal and this plan are the authoritative coordination layer for corrected scope.
+- **Parent should stay open:** no T1750 submission should claim epic completion until the remaining active children are actually closed out.
+
+---
+
+## Immediate next actions
+
+1. Route coders to **T2271**, **T2274**, and **T2275** first.
+2. Treat **T2268** as the merged owner of the canceled **T2272** claimability/CLI work.
+3. Keep **T2270** behind **T2274**.
+4. Do not list **T2272** or **T2273** as active child work in future evidence.

docs/t1750-decomposition-strategy.md

@@ -0,0 +1,52 @@
+# T1750 — Live Decomposition / Coordination State
+
+Updated during the 2026-03-14 heartbeat work session.
+
+## Epic status
+
+T1750 remains **open**. Do not close the epic until every required child is either `done` or explicitly superseded/canceled with the parent state reconciled.
+
+## Current child inventory
+
+| Task | Live status | Role in the epic | Coordination notes |
+|---|---|---|---|
+| T2267 | `blocked` | Historical metadata-inheritance attempt | Blocked earlier because the task was provisioned against the wrong repo/worktree. During this session the child metadata was backfilled to `repo=/home/ubuntu/taskcore`, `parentId=1750`, `reviewer=kelvin` so a future revive/retry will point at the correct repo. Treat T2271 as the forward execution slot for this bug. |
+| T2268 | `analysis.waiting` | Active queue claimability / review-deadlock fix | **Merged scope from canceled T2272 lives here.** This task now owns both dispatch-side guards (`scheduler.ts`, `dispatcher.ts`) and claim-side classifier / CLI listing work (`claimability.js`, queue listing, reviewer deadlock detection). |
+| T2269 | `analysis.waiting` | Active decomposition-state reconciliation fix | Covers idempotent / partial `DecompositionCreated` handling and cleanup of superseded children. |
+| T2270 | `analysis.waiting` | Active attention-alert wiring fix | Depends on T2274 hardening the formatter path first. |
+| T2271 | `analysis.waiting` | Active metadata-inheritance fix | **Full scope, not the earlier narrowed wording.** This child must propagate parent metadata needed for real execution: `repo`, `base_branch`, `informed`, `consulted`, `parentId`, and other operational/custom fields, while letting child-specified values override defaults. |
+| T2272 | `canceled` | Duplicate | Canceled intentionally after its claimability / CLI scope was merged into T2268. Do not treat it as an active child. |
+| T2273 | `done` | Completed decomposition-state slice | Finished. Keep it out of the active-child plan. |
+| T2274 | `analysis.waiting` | Active attention formatter crash fix | Independent Wave 1 work. |
+| T2275 | `analysis.waiting` | Active obsolete / blocked task cleanup tooling | Independent Wave 1 work. |
+
+## Execution waves
+
+### Wave 1 — ready in parallel
+- **T2271** — metadata inheritance (full parent-metadata propagation scope)
+- **T2274** — attention formatter crash hardening
+- **T2275** — obsolete / blocked task cleanup tooling
+
+### Wave 2 — after Wave 1 context is in place
+- **T2268** — merged queue claimability + CLI + review deadlock fix
+- **T2269** — decomposition state reconciliation / idempotent child materialization
+
+### Wave 3 — depends on T2274
+- **T2270** — wire attention output into automated alerting
+
+## Coordination repairs applied in this session
+
+To reconcile live task state with the plan, the following metadata was backfilled onto the outstanding children (`T2267`, `T2268`, `T2269`, `T2270`, `T2271`, `T2274`, `T2275`):
+
+- `parentId=1750` — so child prompts can load parent journal context
+- `reviewer=kelvin` — so review-ready children do not deadlock with a null reviewer
+- `repo=/home/ubuntu/taskcore` — so future code worktrees target the actual taskcore repo instead of the wrong workspace or no repo at all
+
+This does **not** change the immutable task descriptions already stored in state, so the parent journal + this document are the live source of truth for the merged / corrected scope until the children are executed and closed.
+
+## Practical guidance
+
+- If a coder picks up **T2271**, they should implement the **full** metadata propagation bug described in the parent journal, not just priority/assignee/reviewer defaults.
+- If a coder picks up **T2268**, they should treat the canceled **T2272** claimability / CLI scope as in-scope.
+- Do **not** list T2272 or T2273 as active work items in future T1750 submissions.
+- Keep the epic open; this session only repaired coordination / metadata drift and did not complete the children.