retrospective-skill-improver

name	retrospective-skill-improver
description	Retrospectively analyzes target run work traces and improves durable skill packages when reusable guidance, SOPs, examples, or package structure should change.

Retrospective Skill Improver

Purpose

Use work trace evidence for a target run/agent to improve editable skill packages. A useful update turns observed friction into stable guidance that future agents can reuse.

Work trace body entries use role/event labels such as user, assistant, tool, and neutral trace events. The task message and manifest provide the target identity.

Workflow

1. Read the task message for dynamic scope: work trace paths, editable skill roots, package trees, target AgentRun id, and completion message type.
2. Read the relevant work trace files as retrospective Skill Improvement evidence.
3. Inspect each relevant editable skill package, starting from SKILL.md and then any referenced files needed to understand the current guidance.
4. Decide whether the evidence supports a durable improvement:
   • explain that current guidance stands;
   • revise a concise rule, example, or SOP;
   • add or update a reference/template/example file;
   • improve package flow or file organization.
5. Use the listed editable skill roots for file edits.
6. After meaningful durable file changes, send the final skill_update once using the task-supplied completion target.

Evidence Interpretation

Use visible user messages, visible assistant messages, tool calls, tool results/errors, neutral trace events, retries, corrections, and feedback signals. Prefer repeated or explicit evidence over isolated inconvenience.

Base improvements on the observable evidence present in the readable work trace.

High-signal patterns include:

• explicit user correction or future-facing guidance;