name

doom-coordinator

description

Coordinates DOOM codebase exploration in the doom-explore/ directory. Use this skill whenever you are about to dispatch a doom-explorer or doom-synthesizer subagent, when starting a session in doom-explore/, when verifying task completion, or when resuming failed exploration tasks. Handles manifest.json state, reconciliation of stuck tasks, post-dispatch verification, and resume-from-failure flow.

doom-coordinator

You are the coordinator for DOOM codebase exploration. Full design is in SPEC.md. The subagent contracts are in .claude/agents/doom-explorer.md and .claude/agents/doom-synthesizer.md. Your job is to orchestrate them and keep manifest.json consistent.

Hard rules

Never Read or Grep doom-src/ yourself. The subagent does that. Doing it defeats the context-isolation design.
Never dispatch a subagent without following this loop.
Never hand-edit manifest.json status fields. Use coordinator.py.

On session start

python coordinator.py reconcile   # mark stuck 'running' rows as 'failed'
python coordinator.py stuck       # list rows needing re-dispatch

If stuck lists rows, offer the user to resume them before new work.

Before each dispatch

Allocate the next task ID (t-NNNN, zero-padded, monotonic).
Append a row to manifest.json: status: "pending", correct kind (explore or synthesize). For synthesize, fill inputs.
Dispatch via the Agent tool. Project .claude/agents/ types are not registered with the harness, so use subagent_type: general-purpose and tell the agent to follow the appropriate .claude/agents/*.md contract.

After each dispatch

python coordinator.py verify <task_id>

Prints ok if the subagent already flipped the row to complete.
Otherwise marks the row failed with a reason so nothing stays stuck.

Resuming a failed task

Re-dispatch the same subagent type with:

same task_id
resume_from: <artifact path from manifest>

The subagent must append a ## Update <UTC ISO timestamp> section rather than overwrite. That rule lives in the subagent contracts; the coordinator just needs to pass resume_from.

Cost

Subagents use model: claude-haiku-4-5-20251001 (set in each agent's frontmatter — don't override unless the user asks).

doom-coordinator

Hard rules

Never Read or Grep doom-src/ yourself. The subagent does that. Doing it defeats the context-isolation design.

Never dispatch a subagent without following this loop.

Never hand-edit manifest.json status fields. Use coordinator.py.

On session start

python coordinator.py reconcile # mark stuck 'running' rows as 'failed' python coordinator.py stuck # list rows needing re-dispatch

If stuck lists rows, offer the user to resume them before new work.

Before each dispatch

Allocate the next task ID (t-NNNN, zero-padded, monotonic).

Append a row to manifest.json: status: "pending", correct kind (explore or synthesize). For synthesize, fill inputs.

Dispatch via the Agent tool. Project .claude/agents/ types are not registered with the harness, so use subagent_type: general-purpose and tell the agent to follow the appropriate .claude/agents/*.md contract.

After each dispatch

python coordinator.py verify <task_id>

Prints ok if the subagent already flipped the row to complete.

Otherwise marks the row failed with a reason so nothing stays stuck.

Resuming a failed task

Re-dispatch the same subagent type with:

same task_id

resume_from: <artifact path from manifest>

The subagent must append a ## Update <UTC ISO timestamp> section rather than overwrite. That rule lives in the subagent contracts; the coordinator just needs to pass resume_from.

Cost

Subagents use model: claude-haiku-4-5-20251001 (set in each agent's frontmatter — don't override unless the user asks).