[skill-optimizer] Daily Skill Optimizer Improvements - 2026-04-23

[github-actions]

Skill Optimizer Run Summary

• Run mode: dry-run
• Init status: ✅ Success (exit code 0)
• Run status: ❌ Failed (exit code 1)
• Run URL: https://github.com/github/gh-aw/actions/runs/24863909662

Evidence

The optimizer failed in run.log with:

Unhandled error: Error: target.repoPath: target repo has uncommitted changes
(optimize.requireCleanGit is enabled) — Run: git stash or commit your changes

init.log shows the optimizer initialized successfully with:

• Surface: prompt
• Models: claude-sonnet-4.6
• Tasks: up to 10 per run
• Iterations: up to 1
• Target: 80% pass rate

Because run_status: 1, no benchmark or optimization results were produced — the three improvements below address both the immediate failure and preventative quality enhancements.

---

Improvements

1. Add a pre-flight git stash step to the skill-optimizer workflow

Problem: The daily skill-optimizer job fails whenever the workspace contains uncommitted changes (e.g., from a previous make recompile or make init that modifies .skill-optimizer/skill-optimizer.json). The optimize.requireCleanGit guard blocks every run that doesn't start from a perfectly clean tree.

Action: In the .github/workflows skill-optimizer workflow, add a step before invoking the optimizer that runs git stash --include-untracked (and a corresponding git stash pop in a finally block), or configure optimize.requireCleanGit: false in .skill-optimizer/skill-optimizer.json if stashing is too disruptive.

Expected impact: Eliminates the recurring run_status: 1 failure so the optimizer can actually execute dry-runs and benchmarks, enabling continuous skill quality improvements.

---

2. Increase maxTasks and maxIterations in .skill-optimizer/skill-optimizer.json

Problem: answers.json and init.log show the optimizer is configured for at most 10 tasks and 1 iteration. With a single iteration and a 80 % pass-rate target, the optimizer has very little room to explore improvements to SKILL.md before giving up. A single iteration that fails to reach 80 % produces no committed improvement.

Action: Update .skill-optimizer/skill-optimizer.json to increase maxTasks to at least 20 and maxIterations to 3. This gives the optimizer multiple attempts to refine skill guidance before concluding a run, which is especially valuable given that SKILL.md covers many complex topics (MCP configuration, YAML gotchas, Go patterns).

Expected impact: Higher probability of reaching the 80 % pass-rate target in each run, producing meaningful SKILL.md improvements rather than always terminating after a single failed iteration.

---

3. Add task coverage for critical SKILL.md topics that are under-represented

Problem: The current optimizer configuration does not explicitly define task scenarios for several high-risk areas documented in AGENTS.md and SKILL.md: (a) ANSI escape code prevention in YAML files, (b) the needs.activation.outputs.* deprecation and correct steps.sanitized.outputs.* usage, and (c) requireCleanGit-aware CI setup. Without targeted tasks, the optimizer cannot measure whether SKILL.md actually teaches these topics well.

Action: Add at least 3 scenario-based evaluation tasks to the skill-optimizer task suite (in .skill-optimizer/) covering: (1) a prompt asking how to avoid ANSI codes in compiled workflow YAML, (2) a prompt about migrating needs.activation.outputs.text to the new form, and (3) a prompt about configuring the optimizer to run without requiring a clean git state.

Expected impact: Measurable coverage of the most error-prone topics in the codebase, allowing the optimizer to detect and fix gaps in SKILL.md before those gaps cause CI failures or incorrect agent behavior.

Generated by Daily Skill Optimizer Improvements · ● 200.9K · ◷

• expires on Apr 30, 2026, 11:28 PM UTC