[thecolormaroun]

Based on #55 (refactor/david-matt-import-cleanup), and this PR targets that branch. Merge #55 first, then retarget this PR to main (GitHub does it automatically when #55's branch is deleted) and rebase, because #55 is squash-merged. Draft; not merged, not installed.

1. Admission rule

docs/upstream-policy.md. A repo can be a tracked provider ("tracking": "weekly", on the weekly maintenance lane) only if all four hold:

1. It has an OSI license file.
2. It's published for other people: a README written for users, stable skill paths, real commits rather than snapshot dumps.
3. Its maintainer doesn't silently delete skills, or it ships versioned releases.
4. Its skills are general-purpose, not tied to the author's machine, accounts, agents or private repos.

Anything else is "pick-only": Stack imports one reviewed skill, which becomes Stack-owned and stops tracking upstream.

Enforced in code:

• Every provider in registry/upstreams.json has "tracking" (schema updated). scripts/stack-maintenance.py rejects a missing or unknown value.
• The audit never observes a pick-only provider. Its observation row is pick_only, so prepare can't propose it. Supplied refs may include or leave out pick-only providers; a missing weekly one still fails closed.
• scripts/materialize-maintenance-proposal.py and scripts/discover-upstream-updates.py refuse a provider that isn't weekly.
• scripts/import-pinned-skills.py refuses a pick-only provider before any network access. For a weekly one it checks an OSI license id, a regular license file and a README in the exact upstream checkout. Criteria 2–4 need a person, so they live in the audit table.
• tests/test_upstream_admission.py covers all of this and requires every provider to have a row in the audit table.

I didn't edit ~/.claude/scheduled-tasks/stack-upstream-sync/SKILL.md; another session owns it.

2. davidondrej/skills is now pick-only

This reverses part of #55. #55 kept David's 12 surviving skills on the lane. They're now Stack-owned at their last inspected commit (5c99080), the same way #55 converted the others: moved into family roots, names and commands unchanged, origin and MIT notice kept.

david-anti-sleep, david-deepapi, david-fable-safe-prompt, david-read-all-adrs (engineering) · david-browser-harness (qa) · david-cmux, david-codex-subagent, david-goal-loop, david-handoff (orchestration) · david-deep-research (knowledge) · david-effective-agent-skills, david-setup-help (platform)

David has no import rule and no allowlisted root left, and skills/imported/david/ is gone. The provider row stays as a pick-only record. This also clears Monday's blocker: upstream flattened every path on 2026-11-09 (ad5bedd), so the lane would have failed on all 12.

3. Provider audit

Audited at each repo's HEAD on 2026-11-13. Nothing changed except David.

Provider	Repo	Verdict	Fails	Mode now
matt	mattpocock/skills	Pass	None. Removals are in CHANGELOG; 10 releases	weekly
emil	emilkowalski/skills	Pass	None. No releases, but nothing ever removed	weekly
stewdio	StewdioAI/blank-page	Pick-only	3 (ten days old, no releases or track record)	weekly
gstack	garrytan/gstack	Pick-only	4 (63 of 71 SKILL.md files call ~/.claude/skills/gstack/bin)	weekly
compound-engineering	EveryInc/compound-engineering-plugin	Pass	None. 242 releases	weekly
impeccable	pbakaus/impeccable	Pass	None. Apache-2.0	Stack-owned copy
taste-skill	Leonxlnx/taste-skill	Pick-only	2 (layout restructured 4 times in 8 months)	Stack-owned copy
ui	ibelick/ui-skills	Pass	None. 8 releases	not yet a provider
david	davidondrej/skills	Pick-only	2, 3, 4 (every commit is "Publish sanitized skills snapshot"; 18 skills gone without notice; history rewritten; cmux and Pi coupling)	pick-only

Decisions for you, recorded in the policy doc:

• gstack fails only because each skill needs gstack's own public runtime. Accept that kind of dependency and it passes.
• stewdio is worth re-auditing after its first release.
• Upstream merged the impeccable and taste-skill collections into single skills, so Stack's copies are stale relative to them.

4. Shared agent context

• AGENTS.md is the single instruction file. It's config/CLAUDE.md moved with git mv (history kept) plus two sections: shared agent context and upstream skills. CLAUDE.md at the root is a symlink to it. config/CLAUDE.md had no root counterpart and no loader; only a doc test read it, and that test now reads AGENTS.md.
• docs/context/state.md: current state, open PRs, blockers, what's installed. Rewritten each time.
• docs/context/decisions.md: append-only, dated, one line per decision with its PR. Backfilled with fix(runtime): publish the Claude target as a skills-dir plugin #46 (skills-dir plugin runtime), refactor(design): drop Twombly from micro-branch; the local board is the only canvas #52 (Twombly dropped as a paid product), and this PR twice (David pick-only, admission rule).
• docs/context/handoff.md: author, date, done, next.
• The AGENTS.md rule: read state.md and handoff.md before starting; update both before ending a session that changed anything; no secrets or personal data. I added one exemption: scheduled automations that run from dedicated clean clones don't edit these files, because writing them would dirty a checkout that must stay clean. Their receipts are the record.
• tests/test_agent_context.py fails if CLAUDE.md stops being a symlink to AGENTS.md, if any context file is missing, or if the files contain home paths, tokens or email addresses. These files ship with docs/ in the published runtime, and the publisher already rejects /Users/... paths.

Watch out

• The ibelick pin adds provider ibelick. Whichever PR lands second has to give it "tracking": "weekly", or validate_upstream_metadata fails. It passes the rule.
• Expect conflicts in generated files with feat(emil): import new emil skills (gestures, learn-taste) #54 and the ibelick pin. Rebase and regenerate; don't hand-edit them.

Validation

• python3 -m unittest discover -s tests: 549 tests, all pass.
• Registry, classifier and layout checks pass. A classifier rerun is a no-op.

🤖 Generated with Claude Code