Brain Task Review

Task Reconciliation

Read the installed brain skill and verify the intended brain with brain --brain <root> --json doctor and brain --brain <root> path. All task writes use the CLI. The Python 3 stdlib helper below opens SQLite read-only for a complete snapshot and a readback audit; it never mutates tasks or infers closure. Local Brain installs this skill and its audit helper through Settings -> CLI & agents at ~/.agents/skills/brain-task-review. For a manual installation, copy this whole directory, including scripts/, into the agent's skills directory. A manual copy is unmanaged: Settings reports a conflict, refuses to overwrite it, and leaves it in place during uninstall. To switch to app-managed updates, move the manual directory aside first, then install through Settings.

Review The Whole Backlog

At the start of a daily import, snapshot the backlog into that run's scratch directory. Resolve scripts/task_review.py relative to this skill directory:

python3 <skill-dir>/scripts/task_review.py --brain <root> snapshot > tasks-before.json

activeIds contains every open, in-progress, waiting, and blocked task, without a top-N limit. tasks also includes terminal records for duplicate lookup and readback, with task fields, source links, evidence chunks, and stateDigest. Keep these private artifacts inside the brain's import scratch directory.

Before creating an action, compare it against active AND terminal tasks by the underlying obligation, participants, project, source thread, and relevant event, invoice, or billing period. Similar titles alone are not proof of duplication; different invoices, recurring periods, or distinct deliverables remain separate. Repeated reminders should add evidence to the existing task. Do not resurrect completed/cancelled work unless newer evidence clearly creates a fresh obligation.

After importing fresh evidence, review every initially active task and every new task created during the run. This includes undated and waiting tasks, old threads outside the daily source window, and tasks omitted by plan-day limits. Read the original obligation and newer relevant source text, not just task titles or AI summaries. Search across related conversations and records, not only the original thread. Follow linked source identities to live providers when needed.

• Done: evidence establishes the actual requested outcome. A direct external reply can resolve a reply task; a forward to a colleague does not. Sending a reply does not prove payment, delivery, signature, or implementation.
• Cancelled: evidence establishes abandonment, replacement, or that the action's specific opportunity has passed. An arrival instruction for a past stay can expire; an unpaid bill remains actionable after its due date. Record the event/date or decision supporting cancellation. Age alone is insufficient.
• Duplicate: confirm the same obligation, keep one canonical task, add any missing source links/evidence to it, and cancel the others. Put the canonical task ID and consolidation reason in each cancelled task's description.
• Updated: preserve the task's identity while advancing its next action or status. Preserve unrelated description details and links.
• Kept: the inspected evidence still supports the existing obligation.
• Unresolved: evidence or source access is insufficient. Leave it active and record exactly what could not be checked. Do not claim it was fully reviewed.

Use brain --brain <root> --json tasks complete <id> --evidence ... for done and tasks update <id> --status cancelled --description ... --evidence ... for cancellation. Each mutation cites stored source chunks. Use tasks update to attach newer evidence/links to the canonical task. Do not hard-delete tasks. Read back changes before recording their outcome.

Coverage And Readback

Write task-review.json as a JSON array with one entry for every task reviewed:

[
  {
    "taskId": "<id>",
    "outcome": "duplicate",
    "canonicalTaskId": "<kept-task-id>",
    "reason": "Same account, risk scenario and obligation as the canonical task.",
    "checkedSources": ["interaction:<id>#0", "provider query and window checked"],
    "decisionEvidence": ["interaction:<id>#0"],
    "stateDigest": "<digest from a fresh snapshot after CLI writes>"
  }
]

Allowed outcomes are done, cancelled, duplicate, updated, kept, and unresolved. Unresolved entries also require gap. Use actual inspected source refs/queries, not planned searches. Copy each final stateDigest from a fresh snapshot, including for unchanged tasks; do not manufacture it. Every changed task also needs decisionEvidence: source chunk refs in kind:id#index form that support the decision and are attached to that task by the CLI write.

python3 <skill-dir>/scripts/task_review.py --brain <root> snapshot > tasks-after.json
python3 <skill-dir>/scripts/task_review.py --brain <root> audit \
  --before tasks-before.json --ledger task-review.json > task-review-audit.json

The audit re-reads the live database and checks initially active, currently active, and newly created tasks (even if already closed), unique coverage, state digests, outcome/status consistency, and valid canonical tasks. Exit 0 means complete coverage with no unresolved entries; exit 1 means incomplete reconciliation; exit 2 means a helper or input error. Read the JSON even on exit 1. Fix missing/invalid entries; preserve real source gaps. This is a receipt check, not proof that an agent's reasoning is correct. Do not count a blanket unresolved disposition as a reviewed task.

Report active before/after, done/cancelled/consolidated/updated/kept counts, reviewed versus required coverage, and unresolved gaps. If bounded source access prevents full review, say which tasks remain unchecked and why. A successful import audit does not imply successful task reconciliation. Regenerate the daily brief only after this pass so it uses current task state. Follow the existing automation's delivery rules; this skill does not authorize sending messages.