task-supervisor-loop

Use this skill to run a bounded supervisor loop around a task without letting restart docs, project history, and implementation detail accumulate in one prompt.

Inputs

Provide:

• task id or short label
• goal
• done definition
• optional constraints, local paths, non-goals, and write-back targets

If the task spans many docs or files, prepare a pack first with swarm-research-pack.

Workflow

1. Normalize the task into a short task_record with:
   • task_id
   • goal
   • constraints
   • done_definition
   • non_goals
   • write_back_targets
2. Decide whether to build a research_pack.
   • Use swarm-research-pack when the task touches multiple docs, multiple code areas, or ambiguous requirements.
   • Skip the extra phase only when the needed context is already narrow and obvious.
3. Produce an implementation_brief from the task record and research pack.
4. Hand the implementation brief to task-implementer.
5. Hand the implementation result to task-evaluator.
6. If accepted, hand the evaluation result to task-writeback.
7. Keep only the active phase summary and artifact pointers in live context. Do not keep raw phase transcripts resident.

Artifact Rules

Use these artifacts and keep each one short.

• task_record
  • durable task framing
• research_pack
  • selected evidence for this task only
• implementation_brief
  • scope, invariants, files in scope, checks to run
• evaluation_report
  • pass or fail, findings, residual risks, doc mismatches
• restart_note
  • latest outcome, next action, reload set
• plan_delta
  • optional, only when dependency order or critical path changes

Guardrails

• Prefer fresh sessions for research, implementation, evaluation, and write-back.
• Pass structured artifacts, not raw transcripts.
• Expand context only when blocked, then record why the expansion was needed.
• Keep AGENTS.md thin. Do not turn it back into a state ledger.
• Prefer re-derivation from task metadata, diffs, docs, and adjacency over hand-maintained task-specific summaries.
• Spawn subagents only when the user explicitly wants delegation or the task has independent sidecar work.

Output Contract

Use this section order unless the user asks for something else:

# Task Loop
## Task Record
## Research Decision
## Implementation Brief
## Evaluation Gate
## Write-Back Targets
## Next Boundary

Resources

• references/artifact-contracts.md: compact schemas for the handoff artifacts.
• references/session-boundaries.md: what each phase may read, write, and pass forward.
• swarm-research-pack: preferred evidence-pack step before implementation when the task is broad.
• task-implementer: implementation phase skill.
• task-evaluator: evaluation phase skill.
• task-writeback: restart-doc and plan-delta phase skill.