Inspect PR check results and CI failures
Ingest learning candidates from simplify-and-harden-ci
Ingest Handoff blocks from .learnings/HEALS.md (filed by self-healing / self-healing-ci) and surface them as promotion candidates
Deduplicate recurring patterns by stable pattern_key
Emit promotion-ready suggestions for agent context/system prompts

This skill is read-only with respect to the repository (see CI Contract): it does not write .learnings/ entries. Its candidates are emitted as machine-readable output, and promotions are proposed as a PR or comment for human review.

Use self-improvement for interactive/local sessions.

Context Limitation (Important)

CI agents do not have peak task context from the original implementation session. Use this skill to aggregate recurring patterns across runs, not to infer nuanced one-off intent.

Implications:

Favor stable pattern_key recurrence signals over single-run conclusions
Require recurrence thresholds before promotion
Route uncertain or high-impact recommendations to interactive review

Prerequisites

GitHub Actions enabled for the repository
GitHub CLI authenticated (gh auth status)
gh-aw installed for authoring/validation:

gh extension install github/gh-aw

CI Contract

The CI skill must:

Read only PR-scoped data (checks, workflow outcomes, existing learning entries)
Avoid direct code modifications in CI
Emit machine-readable learning output
Recommend promotion only when recurrence thresholds are met

Output Schema

self_improvement_ci:
  source:
    pr_number: 123
    commit_sha: "abc123"
  candidates:
    - pattern_key: "harden.input_validation"
      source: "simplify-and-harden-ci"
      recurrence_count: 3
      first_seen: "2026-02-01"
      last_seen: "2026-02-20"
      severity: "high"
      suggested_rule: "Validate and bound-check external inputs before use."
      promotion_ready: true
  summary:
    candidates_total: 4
    promotion_ready_total: 1
    followup_required: true

Recurrence and Promotion Rules

Track recurrence by pattern_key
Default threshold for promotion:
- recurrence_count >= 3
- seen in >= 2 distinct tasks/runs
- within a 30-day window
Promotion targets:
- CLAUDE.md
- AGENTS.md
- .github/copilot-instructions.md
- SOUL.md / TOOLS.md when using openclaw workspace memory

Authoring Workflow (gh-aw)

Example-only templates live in references/workflow-example.md. Keep examples outside .github/workflows until you explicitly decide to enable CI automation.

When ready:

Copy the template into .github/workflows/self-improvement-ci.md
Customize tool access, outputs, and policy thresholds
Validate:

gh aw compile --validate --strict

Trigger test run manually:

gh aw run self-improvement-ci --push

Heal Handoff Intake

self-healing-ci appends Handoff blocks to .learnings/HEALS.md entries that meet the promotion rule. On each run:

Read .learnings/HEALS.md (read-only) and collect entries with a Handoff block
Map each to a candidate: pattern_key from the HEAL's Pattern-Key, suggested_rule from the Distilled Rule, recurrence fields from the entry metadata
Mark promotion_ready: true when the promotion rule holds, and include the candidate in the output schema alongside simplify-and-harden-ci candidates
Propose the promotion (target file + rule text) as a PR or comment — never write instruction files directly from CI

Integration with Other Skills

Pair with simplify-and-harden-ci to ingest simplify_and_harden.learning_loop.candidates
Pair with self-healing-ci, whose HEALS.md Handoff blocks this skill consumes (see Heal Handoff Intake)
Feed promoted patterns back into self-improvement memory workflow for durable prevention rules

self-improvement-ci

Self-Improvement CI

Install

Purpose