Action Converter

<!-- autoresearch: variation B — sharper output via specificity gates, leverage scoring, banned-phrase lint, open-loop anchoring, empty-state taxonomy -->

${var} — Optional focus area (e.g. health, networking, learning, shipping, crypto, repo). If empty, covers all areas. Treated as a tiebreaker, not a hard filter.

Read memory/MEMORY.md for stated goals, "Next Priorities", tracked items, and current topics. Read the last 7 days of memory/logs/ for recent activity, patterns, and what's already been suggested or done. Read memory/topics/ (every file) for active threads. Read memory/cron-state.json for failing or stuck skills. Read memory/watched-repos.md for repos under attention. Read output/articles/ (last 7 days, filenames only — peek at the 2 most recent for theme). If soul/SOUL.md exists, read it for identity, voice, focus areas. Run gh pr list --state open --limit 20 --json number,title,createdAt,isDraft,reviewDecision,headRefName 2>/dev/null to get open PRs (used to anchor "ship" / "review" / "merge" loops).

Graceful bootstrap — each of the reads above may be missing on cold starts. For every source, if the file/directory is missing or empty (including memory/topics/*.md, memory/cron-state.json, and gh pr list returning empty or erroring), skip it and record BOOTSTRAP: <resource> not yet populated in the run's working notes. Continue with whatever signals are available — the skill must degrade gracefully, never fail. If every single source is empty, fall through to the ACTION_CONVERTER_NO_CONTEXT mode below.

Steps

1. Detect mode

Decide which exit mode this run will produce based on context volume:

• ACTION_CONVERTER_NO_CONTEXT — if BOTH memory/logs/ has 0 entries AND memory/MEMORY.md is the unmodified template (matches "Last consolidated: never" AND "Configure notification channels"). Notify the operator and stop — do not invent actions out of thin air.
• ACTION_CONVERTER_BOOTSTRAP — if memory/logs/ has <3 distinct dates in the last 14 days OR memory/MEMORY.md "Next Priorities" still contains template entries ("Configure notification channels", "Run first digest"). Switch the action pool to setup-completion actions: enable specific skills in aeon.yml, configure missing notification secrets, run the first digest, populate memory/topics/ for the first tracked thread, etc. These are still real, named, completable actions — not generic onboarding advice.
• ACTION_CONVERTER_OK — otherwise. Use the full leverage-scored loop pipeline below.

2. Extract open loops

Build a single deduped list of named open loops from every source above. A loop is a specific in-flight thing, not an area. Each loop captures at minimum: id (short slug), text (one phrase), source (where it came from), age_days, urgency_signal (deadline / blocker / stalled / fresh).

Sources to mine:

• Open PRs — every entry from gh pr list. Loop text: PR #N: <title> with urgency = stalled if >3 days old or review_decision is REQUEST_CHANGES.
• MEMORY.md "Next Priorities" — each bullet becomes a loop. Skip template lines.
• memory/topics/*.md — for each topic file, scan for headings or bullets that look like ongoing work (TODO, WIP, "In progress", "Tracking", trailing question marks, dated items in the last 30 days).
• memory/cron-state.json — every skill with consecutive_failures > 0 OR last_status != success becomes a loop: fix <skill>. Urgency = blocker if consecutive_failures ≥ 3.
• Recent logs (last 7 days) — any line ending in ?, containing "blocked", "next:", "todo", "follow-up", "unfinished", or naming a deferred decision.
• Recent articles (last 7 days) — each new article opens a distribution/syndication loop ("syndicate <slug>") if syndicate-article is enabled, and a feedback loop ("respond to comments on <slug>") if traffic is plausible.
• ${var} — if set, add a synthetic loop "advance ${var}" so at least one action ties to the requested focus area.

Deduplicate by similarity in text. Cap the loop list at 25.

3. Score loops

Score every loop on three 1–5 axes. Total = leverage × urgency × concreteness.

Drop any loop scoring <8 from the candidate pool. If ${var} is set, give a +0.5 leverage bump to loops touching that area.