Learning Aggregator Ci

Learning Aggregator CI

Install

bashCopy code
gh skill install pskoett/pskoett-skills learning-aggregator-ci

For interactive sessions, use:

bashCopy code
gh skill install pskoett/pskoett-skills learning-aggregator

Fallback using the Agent Skills CLI:

bashCopy code
npx skills add pskoett/pskoett-skills/skills/learning-aggregator-ci
npx skills add pskoett/pskoett-skills/skills/learning-aggregator

Purpose

Runs the outer loop's inspect step in CI. Reads accumulated .learnings/ files, groups entries by pattern_key, computes cross-session recurrence, and produces a ranked gap report — all without human interaction.

The interactive learning-aggregator skill is designed for in-session use where the user can review and act on findings immediately. This CI variant runs on a schedule (weekly, per-sprint, or on-demand) and posts its findings as a GitHub issue comment for async review.

Context Limitation (Important)

CI agents do not have session context. They cannot see what the user is currently working on or what task area is relevant. The CI variant scans all .learnings/ entries without relevance filtering. The gap report is comprehensive rather than targeted.

Prerequisites

• GitHub Actions enabled on the repository
• gh CLI authenticated with repo access
• gh-aw extension installed (gh extension install github/gh-aw, v0.40.1+)
• .learnings/ directory with structured entries from self-improvement

CI Contract

Hard rules for headless execution:

1. Read-only — do not modify .learnings/ files, project instruction files (CLAUDE.md, AGENTS.md, .github/copilot-instructions.md), or any repo files
2. Headless — no interactive prompts, no approval gates
3. Structured output — emit findings as YAML under learning_aggregator_ci key
4. Single comment — post one consolidated comment per run, not per finding
5. Deterministic — same .learnings/ state produces the same gap report

Authoring Workflow (gh-aw)

1. Copy references/workflow-example.md into .github/workflows/learning-aggregator-ci.md
2. Customize the schedule for your cadence (supports fuzzy schedules like weekly on mondays)
3. Validate: gh aw compile (optionally add --actionlint --zizmor for full security scan)
4. Push to enable

Persistence and Chaining

• cache-memory: stores aggregation state (pattern groups, recurrence counts) across runs. Survives up to 90 days in Actions cache. Avoids re-scanning unchanged entries on every run.
• call-workflow: triggers eval-creator-ci after aggregation completes to create evals from newly promoted patterns. Compile-time fan-out with proper dependency wiring.
• upload-artifact: persists the gap report YAML for consumption by downstream workflows or human review.

Cache state must declare aggregation schema provenance-v1 and retain canonical occurrence fingerprints, stable task lineage, and terminal-event boundaries. Ignore and rebuild any cache that omits this version or uses an older aggregation schema; aggregate counts from the pre-deduplication contract are not a valid baseline.

Workflow Rules

The CI agent follows these rules in order:

1. Read all files in .learnings/: LEARNINGS.md, ERRORS.md, FEATURE_REQUESTS.md, HEALS.md
2. Parse each entry's metadata: Pattern-Key, Recurrence-Count, First-Seen, Last-Seen, Priority, Status, Area, Related Files, Tags, and optional provenance fields Task-ID, Session-ID, Occurrence-ID, Source-Ref, Copied-From. For HEAL entries, also parse Trigger, Active-Context, and any Handoff block
3. Before grouping, collapse copies with the same entry ID/content or occurrence ID across repo locations, mirrors, forks, forwards, and cloud/local sources. When explicit occurrence IDs are absent, use task/session/source lineage and normalized evidence. Different paths are not independent evidence
4. Group canonical occurrences by Pattern-Key (exact match only — no fuzzy grouping in CI)
5. For each group: count deduplicated recurrences, count distinct tasks from stable provenance, compute the time window, and collect evidence. A legacy entry without stable task/session lineage contributes its declared recurrence once but all unknown-lineage evidence counts as at most one distinct task
6. Flag entries without Pattern-Key as ungrouped
7. Treat promoted, promoted_to_skill, resolved, and wont_fix as terminal for their recorded occurrence. Keep terminal-only groups as history, not promotion candidates. Reopen only for newer active evidence after the latest terminal event; a prior Handoff alone does not re-promote the pattern
8. Classify each actionable group's gap type: knowledge gap, tool gap, skill gap, ambiguity, or reasoning failure
9. Rank groups by: promotion-ready first, then approaching threshold, then by priority (critical > high > medium > low)
10. Emit structured YAML under key learning_aggregator_ci
11. Post gap report as a comment on the triggering issue or as a new issue if running on schedule
12. Do not modify repository files

Promotion threshold (same rule as learning-aggregator and self-improvement): a group is promotion-ready when it has >= 3 deduplicated recurrences, seen in >= 2 distinct tasks proven by stable provenance, within a 30-day window.

Output Schema

yamlCopy code
learning_aggregator_ci:
  version: "0.1.0"
  source:
    run_id: "<workflow run ID>"
    trigger: "schedule | workflow_dispatch | issue_comment"
    scan_date: "YYYY-MM-DD"
  scan:
    entries_total: 42
    entries_with_pattern_key: 35
    entries_ungrouped: 7
    patterns_found: 18
    promotion_ready: 3
    approaching_threshold: 5
  promotion_ready:
    - pattern_key: "harden.input_validation"
      recurrence_count: 5
      distinct_tasks: 3
      window_days: 21
      priority: "high"
      gap_type: "knowledge_gap"
      area: "backend"
      evidence:
        - "LRN-20260301-001: Missing bounds check on pagination params"
        - "ERR-20260308-002: Unconstrained string length caused OOM"
        - "LRN-20260315-003: API params not validated before DB query"
      recommended_action: "Add to project instruction files: Always validate and bound-check external inputs before use"
      eval_candidate: true
  approaching:
    - pattern_key: "simplify.dead_code"
      recurrence_count: 2
      distinct_tasks: 1
      priority: "low"
      needs: "1 more distinct task"
  ungrouped:
    - id: "LRN-20260320-005"
      summary: "Discovered undocumented rate limit on external API"
      recommendation: "Assign pattern_key for future tracking"
  stale:
    - pattern_key: "harden.error_handling"
      last_seen: "2025-12-01"
      recommendation: "Dismiss — not seen in 90+ days"
  summary:
    promotion_ready_total: 3
    approaching_total: 5
    ungrouped_total: 7
    stale_total: 1
    followup_required: true

Recommended Outputs

Trigger Configuration

Recommended: weekly schedule + manual dispatch

yamlCopy code
on:
  schedule:
    - cron: '0 9 * * 1'  # Monday 9am UTC
  workflow_dispatch:
  issue_comment:
    types: [created]

The schedule ensures regular outer-loop cadence. Manual dispatch allows on-demand runs after incidents or sprints. Issue comment trigger allows /aggregate-learnings commands.

Integration with Other Skills

Upstream (feeds from)

• self-improvement (interactive) — produces .learnings/LEARNINGS.md, ERRORS.md, FEATURE_REQUESTS.md entries
• self-healing / self-healing-ci — produce .learnings/HEALS.md entries including Handoff blocks
• self-improvement-ci — emits learning candidates as machine-readable output (artifacts/comments); it is read-only and does not write .learnings/ files itself
• simplify-and-harden-ci — produces learning_loop.candidates consumed by self-improvement-ci

Downstream (feeds into)

• harness-updater (interactive) — takes promotion-ready patterns from the gap report and applies them
• eval-creator-ci — takes eval candidates and creates permanent test cases
• Human review — gap report posted as issue comment for team triage

Data Flow

Copy code
self-improvement → .learnings/*.md   ←  self-healing(-ci) → HEALS.md
                       ↓
              learning-aggregator-ci (scheduled)
                       ↓
              gap report (issue comment + artifact)
                       ↓
              harness-updater (interactive, human-gated)
                       ↓
              eval-creator-ci (creates evals from promoted patterns)

Differences from Interactive Version