| name | retrospective |
| description | Use when a session ends, checks fail, user corrections arrive, or AGENTS/skills/process instructions may need improvement. |
Retrospective Skill
Goal
Improve the agent system without bloating always-on context.
When a recurring workflow branches, prefer recording a bounded child-worker
handoff and the Spark lane cue rather than flattening the lesson into a wider
local pass.
Process
- Summarize outcome.
- Identify friction and root causes.
- Classify update target: AGENTS.md, nested override, skill, PRD/ADR, Task Master, CI/tooling, or no action.
- Prefer skill updates over root AGENTS.md updates.
- Prefer automated checks over prose.
- Return proposed patches separately.
Durable lessons
- If the friction is repeated root-level replanning, over-broad ownership, or
a missing Spark lane cue, prefer the narrowest routing surface first:
coordinator TOML for delegation, implementer TOML for slice boundaries, or
code-intelligence for tool-ladder changes.
- If the lesson is truly recurring and global, promote it out of this skill;
otherwise keep it here, in Task Master notes, or in the narrowest durable
repo-local skill or agent surface.
- If the same lesson is likely to recur, prefer a tracked skill, agent TOML,
or docs update before writing anything to memory.
- If a branch is rewritten, force-pushed, or replaced, refresh live PR checks
and mergeability evidence against the current head before treating older
results as truth.
Output
- Outcome:
- Friction:
- Root cause:
- Proposed improvement:
- Recommended file:
- Proposed patch:
- Do not change: