task-handoff

Tier-A structured agent handoff — no new infrastructure, just discipline with the task store. The task store is the coordination layer; conversation context is not.

Protocol

When Agent A completes work that unblocks Agent B:

1. Agent A writes its outcome to the task store — status, findings, any blockers, next steps
2. Agent A creates Agent B's task in the task store with assigned_agent set
3. Agent A marks its own task done and records the handoff ID in notes
4. Agent B reads from the task store to find its assignment — not from memory or conversation

Example: Feature implementation → QA review

# Agent A (dev): marks feature done, creates QA task
ts-create "myproject" "qa-review-api" "QA the REST API implementation" "phase1" "high"
# -> returns downstream_id

# Agent A sets blocked_by to empty (upstream gates were cleared)
sqlite3 ~/.claude/task-store.db "UPDATE tasks SET blocked_by='[]' WHERE id='$downstream_id'"

# Agent A marks own task done
ts-status "$upstream_id" "done"

# Agent B (QA): reads task store for its assignment
ts-project "myproject" | grep "qa-review-api"
# -> finds task, starts work

# After QA, Agent B writes gate verdict
ts-gate "$feature_task_id" "passed" "qa-agent"

Key Rules

• Task store is the source of truth — not conversation, not memory
• blocked status is a hard gate — never advance to in_progress while blocked
• Handoffs are explicit — always write to task store before spawning downstream
• Failure propagates — if Agent B fails after retries, escalate to PD/council

Anti-patterns

• Spawning an agent without checking the task store first
• Storing state in conversation context (it gets lost between sessions)
• Advancing a blocked task to in_progress
• Skipping gate verification for "quick" downstream tasks

Relationship to Tier-B (future)

Tier-B adds a file-based FIFO queue (~/.claude/agency-rooms/{room}/queue/) so agents can poll asynchronously. Tier A is the prerequisite — implement this first to establish the discipline.