Self Improve

${var} — Mode selector, optionally with a focus area, as mode or mode:focus.

• empty or improve → improve mode: find and fix the highest-impact issue from recent logs, then propose + apply the fix via PR (default).
• improve:<area> (or a bare area like notifications) → improve mode focused on that specific area (e.g. heartbeat, notifications, memory).
• audit → audit mode: review what the agent did, what failed, and what to improve; save a full review and apply safe, obvious fixes directly.
• audit:<area> → audit mode focused on that specific area (e.g. reliability, memory).

Setup (both modes)

Parse ${var} into a mode and an optional focus area:

• Split on the first : — the part before is the mode, the part after is the focus.
• If the mode is audit → run the Mode: audit branch below (focus = optional area to concentrate the review on).
• If the mode is improve or empty → run the Mode: improve branch below (focus = optional area to fix).
• If the token is neither keyword but non-empty (e.g. notifications) → treat it as improve mode with the whole ${var} as the focus area (backward compatibility).

Then:

• Read memory/MEMORY.md for high-level context and goals.
• Read recent memory/logs/ (improve mode: last 2 days; audit mode: last 7 days) for errors, failures, and quality issues.

If a focus area is set, concentrate the run on that area.

Mode: improve (default)

Improve the agent itself based on recent performance. ONE change per run.

Steps

1. Check for open improvement PRs — don't pile up unreviewed work:

   OPEN_PRS=$(gh pr list --state open --json title,number --jq '[.[] | select(.title | test("^(fix|feat|chore)\\("; "i"))] | length')

   If there are already 3+ open improvement PRs, log "self-improve: 3+ open PRs, waiting for review" and exit. Don't create more debt.

2. Identify what to improve. If the focus area is empty, scan for issues:

   • Read memory/logs/ from last 2 days — look for:
     • Skills that failed or produced low-quality output
     • Errors, timeouts, "zero output", rate limiting
     • Notifications that didn't send or were truncated
     • Memory consolidation problems
   • Read memory/cron-state.json for skills with low success rates
   • Read output/articles/repo-actions-*.md from last 7 days for self-improvement ideas
   • Pick the highest-impact, smallest-effort fix. One change per run.

3. Understand the area you're fixing. Read the relevant files:

   • Skills: skills/{name}/SKILL.md
   • Config: aeon.yml
   • Workflows: .github/workflows/*.yml
   • Agent instructions: CLAUDE.md
   • Dashboard: apps/dashboard/ (if UI-related)

   Understand the current behavior before changing anything.

4. Implement the fix. Make minimal, targeted changes:

   • If a skill prompt is unclear → rewrite the ambiguous section
   • If a skill is hitting rate limits → add backoff logic or reduce frequency
   • If output quality is low → tighten the prompt, add examples, clarify format
   • If a notification is broken → fix the formatting or truncation
   • If a config is wrong → fix aeon.yml

   Do NOT:

   • Rewrite entire skills from scratch
   • Add new features (that's create-skill's job)
   • Change the core architecture
   • Modify secrets or environment variables

4b. Dry-run gate. Before opening the PR, execute the improved skill once with synthetic secrets, so a regression never reaches production having run only with real credentials. Let $skill be the skill you edited:

DRYRUN_VERDICT="output/.dry-run/$skill.json" bash scripts/dry-run.sh run "$skill" || true

Read output/.dry-run/$skill.json: passed: true (or skipped: true, when the SKILL_DRYRUN repo variable is 0) continues. passed: false means revert the edit and stop (log self-improve: dry-run gate failed for $skill with the verdict reasons[]; do not open the PR). Put the verdict under a ## Dry-run section in the PR body. The gate is structural (exit, output, declared mode, declared requires:); no real credential enters the run.

5. Create a branch and PR:

   git checkout -b fix/self-improve-${today}
   git add -A
   git commit -m "fix: [description of what was improved]
   
   Problem: [what was failing/degraded]
   Fix: [what was changed]
   Evidence: [log entries, error messages, success rates]"

   Open a PR:

   gh pr create --title "fix: [short description]" \
     --body "## Problem
   [What was failing or degraded — cite specific log entries or error messages]
   
   ## Fix
   [What was changed and why]
   
   ## Evidence
   - [Relevant log entries]
   - [Success rate before: X%]
   - [Error pattern: ...]"

6. Notify. Send via ./notify:

   self-improve: [what was fixed] — PR: [url]

7. Log (see the shared ## Log section below).