Deep plan ingest

Deep Plan Ingest

A plan document is a snapshot; it goes stale the day implementation starts. This skill moves the verified knowledge out of the plan and into the three files future agents actually load — so the research paid for once keeps paying.

Target	Purpose	Audience
AGENTS.md	Operating rules + ground-truth facts for agents working in this repo	Every future agent session
docs/ARCHITECTURE.md	Component map, data flow, framework mechanisms — with citations	Agents and humans exploring the system
ROADMAP.md	The phased roadmap with live phase status	Anyone asking "what's next?"

The Prime Directive (inherited from deep-plan)

Only VERIFIED knowledge gets ingested. Stale evidence is re-verified or dropped.

• Claims tagged UNVERIFIED in the plan never enter AGENTS.md or docs/ARCHITECTURE.md. They may appear in ROADMAP.md only as explicit open questions.
• CORRECTED entries win over the beliefs they corrected — ingest the correction, never the original mistake.
• A citation is copied along with its claim. A fact that loses its file:line / vendor path / doc URL on the way over was not ingested, it was laundered.

Progress checklist

Copy this into your response and check items off:

Ingest Progress:
- [ ] Step 1: Locate the plan and confirm targets
- [ ] Step 2: Staleness check — re-verify claims touched since the plan was written
- [ ] Step 3: Ingest AGENTS.md (merge, don't clobber)
- [ ] Step 4: Ingest docs/ARCHITECTURE.md (merge, don't clobber)
- [ ] Step 5: Ingest ROADMAP.md with phase status
- [ ] Step 6: Report — written / updated / skipped / dropped-as-stale

Step 1 — Locate the plan and confirm targets

• Plan path: use the path given with the invocation ($ARGUMENTS on platforms that substitute it); else try docs/PLAN.md, then docs/*PLAN*.md. If several candidates exist, ask ONE question with your recommended pick.
• Target detection: if the repo already uses CLAUDE.md or AGENT.md as its agent-instructions file, merge into that file instead of creating a parallel AGENTS.md. Never create a second competing instructions file.
• Read the entire plan before writing anything.

Step 2 — Staleness check

The plan was true when written; the repo has moved. Before ingesting:

1. git log --oneline <plan-file-mtime-or-date>..HEAD — list commits since the plan (or since its "Generated on" date).
2. For every plan claim whose cited file appears in those commits, re-open the file and re-verify before carrying the claim over.
3. Claims that no longer hold are dropped and listed in the Step 6 report — never silently ingested.

Skip nothing here: ingesting a stale "fact" into AGENTS.md poisons every future session that loads it.

Step 3–5 — Write the three files

Follow the exact structures in the references (read the one you need before writing that file):

• AGENTS.md → references/agents-template.md
• docs/ARCHITECTURE.md → references/architecture-template.md
• ROADMAP.md → references/roadmap-template.md

Merge policy (applies to all three):

• If the target does not exist, create it from the template.
• If it exists, update it section by section: refresh sections this skill owns, and leave human-authored sections untouched. When ownership is unclear, append under a clearly marked section rather than editing prose you did not write.
• Keep it lean. AGENTS.md is loaded into context every session — hard budget under 150 lines; move depth to docs/ARCHITECTURE.md and link to it. Prefer deleting an outdated line over adding a clarifying one.
• No duplication between the three files: rules live in AGENTS.md, structure in ARCHITECTURE.md, sequencing in ROADMAP.md. Cross-link instead of copying.

Before reporting, run the mechanical checks — commands, not judgment calls; all must pass:

1. wc -l on the AGENTS-type file → 150 or less; over budget = cut lines until it passes, never present-and-apologize.
2. Repo metadata (remotes, hosting, author) must match git remote -v / git config output exactly — never construct a GitHub/GitLab URL from the repo's name; no remote = say so.
3. Every claim carried over kept its citation — grep the outputs for bare assertions that lost their file:line on the way.

Step 6 — Report

End with a concise report:

• Written/updated: each file with a one-line summary of what changed.
• Dropped as stale: each plan claim that failed re-verification, with the commit or file that invalidated it.
• Skipped: UNVERIFIED items excluded (count is enough).
• Suggested follow-up: if many claims were stale, recommend re-running deep-plan for the affected area instead of patching the docs by hand.