Action Converter

${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 ") if syndicate-article is enabled, and a feedback loop ("respond to comments on ") 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.

4. Convert loops to actions

Convert the top loops into actions until you have 5 distinct ones. Constraints on every action:

1. Specificity gate — must name at least one of: a file path, a PR number, a person/handle, a project/repo, a tool/CLI command, a URL, a tracked entity from MEMORY.md. Generic "reach out to people" / "review your goals" / "explore opportunities" fails this gate.
2. Banned-phrase lint — reject any action whose action text contains: go for a walk, drink water, take a break, reflect, journal, meditate, brainstorm, review your, think about, consider, look into, explore opportunities, reach out to people, network with, clean up your inbox, organize your, plan tomorrow, do some reading, check social media. These are filler, not actions.
3. Time estimate — must fit in ≤2 hours; bias toward 30–60 min slots.
4. Definition of done — one observable check. "PR opened" / "commit pushed" / "message sent to " / "doc has section X with ≥3 items". Not "feel better" or "have more clarity".
5. Anti-template (14-day novelty check) — for each candidate action, extract the verb + main noun. Reject if the same verb+noun appears in any memory/logs/*.md from the last 14 days. (Different verb+same noun is fine — only the bigram blocks.)
6. Score ≥4 on the 1–5 quality scale below. Anything <4 is dropped and replaced from the next loop in the queue.

Quality 1–5: 1 = filler, 2 = vague, 3 = specific but low-leverage, 4 = specific + tied to a real loop, 5 = specific + high-leverage + would visibly move the project today.

If after exhausting the loop list you have <5 surviving actions, fill the rest from the category pool below but only with category-specific candidates that pass all gates above. Categories exist as a fallback, not a checklist:

• Build — ship, write, create, deploy, fix, prototype, refactor against a named file/PR
• Connect — DM/reply/quote a named handle about a named topic; comment on a named PR/issue
• Learn — read a named paper/doc/repo and write a 5-bullet takeaway to memory/topics/
• Health/Energy — only if tied to a named, novel, non-banned action (rare; usually skipped)
• Money — concrete revenue/funding/deal step naming a counterparty
• Position — write a named tweet/cast/post on a named claim
• Explore — lateral move tied to a named external signal from this week's logs

If even with the category pool you can't reach 5, output fewer (3 or 4) and flag ACTION_CONVERTER_THIN in the notify — don't pad.

5. Compose the output

Build one today's shape line: ≤14 words capturing the dominant theme of the 5 actions ("Ship 2 PRs, unblock failing skill, syndicate yesterday's article" — not "be productive today"). This becomes the lede.

Order the 5 actions by descending quality score, then by descending urgency.

6. Send via ./notify

Use this exact format. Telegram-MD friendly. No leading spaces on any line (Telegram renders indents as code blocks).

*5 Actions — ${today}*
Shape: <today's shape line>

1. <action — one imperative sentence, names a specific entity>
why: <≤18 words, what makes this leverage today, names a specific signal>
done: <one observable check>
loop: <loop id or "category:<name>" if filled from pool>

2. <action>
why: <…>
done: <…>
loop: <…>

3. <action>
why: <…>
done: <…>
loop: <…>

4. <action>
why: <…>
done: <…>
loop: <…>

5. <action>
why: <…>
done: <…>
loop: <…>

sources: memory=<lines> logs=<days> topics=<files> prs=<open> cron_failing=<n> mode=<OK|BOOTSTRAP|THIN>

Notification rules:

• Drop any action whose done: line couldn't be written without hand-waving.
• If mode is ACTION_CONVERTER_NO_CONTEXT, skip the action list entirely and notify: *Action Converter — no context yet* plus a one-line pointer ("Populate memory/MEMORY.md or run a skill to seed memory/logs/").
• If mode is ACTION_CONVERTER_BOOTSTRAP, prefix the shape line with Bootstrap mode: and pull all actions from the setup-completion pool.

7. Log to memory/logs/${today}.md

Append:

## Action Converter
- **Mode:** OK | BOOTSTRAP | THIN | NO_CONTEXT
- **Focus:** <var or "general">
- **Shape:** <today's shape line>
- **Actions:** N (quality avg <x.x>/5)
- **Loops anchored:** <list of loop ids surfaced>
- **Loops carried over:** <list of high-score loops not chosen, for tomorrow>
- **Notification sent:** yes

Carrying loops forward in the log is what powers the 14-day novelty check and lets the next run see what's been deferred.

Sandbox note

gh pr list works in the GitHub Actions sandbox via the gh CLI (handles auth internally). If gh is unavailable or returns empty, treat the open-PR loop source as prs=0 and continue — do not block the whole run.

No outbound HTTP is required. All inputs are local files and gh. No new env vars.