auto-workflow

<!-- autoresearch: variation B — sharper output (priority tiers + data-verification gates + delta-against-existing + exit taxonomy) + slug-enable execution branch (validate → commit → PR) -->
> **${var}** — selects the mode:
> - **Analyze (default):** a URL to analyze (GitHub repo, X account, blog, project site, API docs, etc.). Multiple URLs comma-separated. Prefix a URL with `force:` to re-analyze one already in the ledger. Produces a tiered recommendation article + an `aeon.yml` diff — it does **not** mutate `aeon.yml`.
> - **Enable:** `enable:slug1,slug2,…` — flip those skills' `enabled: false → true` in `aeon.yml`, validate each against `skills/`, then commit + open a PR. `enable:dry-run:slug1,slug2` validates and reports without editing, committing, or opening a PR.
>
> Example values: `https://example.com/blog` · `@vitalikbuterin, github.com/foundry-rs/foundry` · `force:https://mirror.xyz/somedao` · `enable:rss-digest,github-monitor` · `enable:dry-run:price-alert`

## Overview

One skill, two ends of the same loop: **analyze** decides *what to enable* for a new watch target; **enable** actually flips the switch. Dispatch `enable:` with the slugs the analyze run put in its MUST tier and you close the loop — recommendation to merged PR — without a second skill.

**Analyze mode** verifies every recommendation is backed by an *observed* signal on the URL, tiers output into **MUST** (2–3 max) / **SHOULD** / **NICE** with a one-line concrete "why", emits a *delta* against the current `aeon.yml` rather than a full config dump, stays silent when existing config already covers the URL, and anchors skill names in `skills.json` (authoritative), not a stale mapping table. It writes an article + updates a ledger; it never edits `aeon.yml`.

**Enable mode** does the mechanical part analyze deliberately leaves to the operator: a slug-scoped `enabled: false → true` substitution in `aeon.yml`, gated by directory presence / current-disabled-state / chain-conflict checks, committed on a fresh branch and shipped as a PR with per-skill rationale. Explicit opt-in only — the operator names the slugs; nothing flips on `main` until they click merge.

---

## Shared preamble (run for both modes)

1. Read `memory/MEMORY.md` for high-level context and skim the last ~3 days of `memory/logs/` — drop anything already reported so you don't re-emit the same signal.
2. Parse `${var}` to pick the branch:
   - `${var}` is empty → **Analyze** branch, empty-input path → exit `AUTO_WORKFLOW_EMPTY`, notify `auto-workflow: set var= to one or more URLs (comma-separated), or enable:slug1,slug2 to flip skills on`.
   - `${var}` starts with `enable:` (case-insensitive) → **Enable** branch. Strip the `enable:` prefix; the remainder is the slug list (which may itself begin with `dry-run:`). Go to **Mode B**.
   - Otherwise → **Analyze** branch. Go to **Mode A**.

---

## Mode A — Analyze: generate recommendations + aeon.yml diff (default)

### A0. Parse input and load context

If the (post-preamble) input is empty → exit `AUTO_WORKFLOW_EMPTY`, notify as above.

Parse `${var}`:
- Split on `,`, trim each entry
- Detect `force:` prefix on any entry → sets `force=true` for that URL (skip ledger dedup)
- Normalize each URL:
  - Add `https://` if scheme missing
  - `twitter.com/` → `x.com/`
  - `@handle` → `https://x.com/handle`
  - Strip trailing `/`, fragment, and tracking params (`utm_*`, `ref`, `src`, `s`, `t`)
  - Strip trailing `.git` on github URLs
- Reject `javascript:`, `data:`, local file URLs → exit `AUTO_WORKFLOW_ERROR` with the bad URL

Read context:
- `memory/MEMORY.md` — operator interests
- `aeon.yml` — CURRENT skill enablement, `var`, `schedule`, `model` per skill (this is the comparison baseline)
- `skills.json` — authoritative installed-skill list
- `memory/topics/auto-workflow-analyzed.md` (if exists) — for ledger dedup

**Ledger dedup:** If a URL is in the ledger with `analyzed_at` within the last 14 days and `force` is not set for it, skip it with `already_analyzed` reason. If ALL inputs are dedup-skipped → exit `AUTO_WORKFLOW_NO_CHANGE`, notify nothing, log a one-line skip entry.

---

### A1. Fetch and classify

For each remaining URL, `WebFetch` with prompt: "Return page title, meta description, all <link rel='alternate'>, og:* meta tags, social handle links (x.com, github.com, t.me, discord), detected RSS/Atom feed URLs, and any token contract addresses (0x… or Solana base58 near the words 'token'/'contract'/'mint'). Report the most recent date on the page. Report the tech stack (Jekyll/Hugo/Next.js/WordPress etc)."

If fetch fails or returns <300 chars of meaningful content, try fallbacks: `/robots.txt`, `/sitemap.xml`, `gh api` for github URLs. If all fail → mark this URL `FETCH_FAILED` with reason and continue to next URL.

Classify into ONE primary category: `github-repo` / `github-org` / `x-account` / `blog-or-news` / `crypto-project` / `api-or-docs` / `research` / `product` / `community` / `personal-site` / `other`.

Extract **concrete signals** (the "why" anchors for later recommendations):
- `feed_urls`: list of RSS/Atom URLs discovered
- `x_handles`: list of X handles linked from page
- `github_repos`: list of owner/repo from page links
- `token_contracts`: list of (chain, address, symbol) tuples
- `last_update`: most recent date found (ISO)
- `update_cadence`: estimate — `active` (<7d old), `steady` (<30d), `quiet` (<90d), `dormant` (≥90d)
- `tech`: stack hint if any

If classification confidence is low (sparse signals, no category clearly matches), mark `UNCLASSIFIED` for this URL and skip to next.

---

### A2. Match signals to installed skills

For each URL, generate candidate skills by intersecting:
- URL `category` and extracted signals
- Skills present in `skills.json`

Use this hint table — but **only emit skills whose slug exists in `skills.json`** (drop any slug not found):

| Category | Hint skills | Requires signal |
|----------|-------------|----------------|
| github-repo | github-mon