fix: restore portable skill contracts

[3metaJun]

Problem

The Windows model checker resolved its default path to a literal ~\\.config\\mstack\\models.json, setup documented only four roles although seven are configured, and several skills still depended on Cursor or legacy worker names. Canonical skill bodies also had no reviewed baseline, so a thinned file could pass inventory checks.

Changes

• Resolve the default model file from the platform home directory and accept ~\\... paths on Windows.
• Document all seven model roles in /setup-mstack.
• Replace legacy Comment Sicko, Task subagent, and built-in create-skill references with portable delegation and skill-authoring guidance.
• Record canonical skill source and reviewed target body digests; strict upstream checks now detect missing, changed, or truncated bodies while preserving adapted skills.
• Add regression coverage and document the new upstream integrity behavior.

Validation

• npm test (78 passed, 1 expected platform skip)
• npm run check-package
• node scripts/check-upstream.mjs --source G:\\agents_temp\\mstack-pstack-full-20260911\\pstack --strict
• git diff --check

npm publishing is intentionally deferred until the remaining fixes are complete.

---

Agent: GPT-6 via Codex

[3metaJun]

Adversarial Review (4 parallel reviewers)

Reviewed dcf8bfa via four independent adversarial passes — code correctness & path handling, test quality & validation re-run, docs/claims consistency, and manifest/digest integrity. Each re-ran the claimed validations and independently recomputed digests.

Overall verdict: APPROVE-WITH-NITS. No blockers; the mechanism works and the PR is internally consistent. The P2 items below directly undercut the two headline claims and are worth a fast-follow.

Verified sound

• Validation claims reproduce exactly: npm test → 79 tests, 78 pass / 1 skip (the skip is the pre-existing Unix-only worktree-audit test, unrelated to this PR); npm run check-package passes.
• All 47 target digests in profiles/upstream-manifest.json match independent recomputation — including the six SKILL.md bodies edited in this very PR (re-pinned consistently).
• Mutation testing confirms strict detection fires on changed, truncated, missing-entry, and stale-entry bodies; CRLF-normalized hashing means Windows autocrlf checkouts won't false-positive; strict provenance (clean checkout, HEAD == pinned commit) fails closed.
• The default-path fix works (literal ~\... bug gone); expandHome handles ~\, mixed separators, UNC passthrough, and correctly leaves ~~/~user literal.
• The seven model roles in /setup-mstack match DEFAULT_ROLES exactly; Comment Sicko / Task subagent / worker type / create-skill are purged from shipped content (remaining hits are the upstream rename map and test guards, by design).
• CI exercises check-upstream --strict against the pinned upstream, so source-side digests are actually enforced, not aspirational.

P2 findings

1. Strict mode fails open when canonicalSkills is absent — scripts/check-upstream.mjs:337-339. Deleting the manifest block (or corrupting it to a non-object) yields only a console.warn and exit 0 under --strict, while the README promises strict fails when canonical bodies are "missing, differs". Aggravated by scripts/sync-upstream.mjs:459, which only carries canonicalSkills forward from the previous manifest and never generates it: a fresh --apply without a readable predecessor produces a baseline-less manifest and the strict guarantee silently evaporates. Suggest artifactProblems += 1 in that branch when strict.

2. The ~\ fix covers only 1 of 3 scripts sharing the same helper — scripts/run-role.mjs:27-29 and scripts/smoke-harnesses.mjs:19-21 still expand only ~/. On Windows, run-role --file ~\.config\mstack\models.json resolves to a nonexistent cwd-relative path, and readModelConfig silently returns defaults — the exact silent-misconfiguration class this PR fixes, one script over.

3. Zero automated coverage of the new canonical check block — the ~35-line block in check-upstream.mjs is exercised only target-side via skill-integrity.test.mjs; every sync-upstream.test.mjs fixture omits canonicalSkills, so CI only ever runs the warn path of the new code. The ~\ branch of expandHome itself is also untested (only the default-path regression is).

4. Unhandled ENOENT when a manifest entry's upstream source SKILL.md is absent — scripts/check-upstream.mjs:346-352. The target side is guarded with existsSync; the source side isn't. Fails closed, but aborts the whole remaining report with a raw stack trace instead of a per-skill diagnostic.

5. README overclaims --apply maintenance, and reflect points at a nonexistent workflow — README.md:307-309 implies the documented workflow maintains canonicalSkills, but nothing generates or re-pins those digests; every reviewed adaptation requires an undocumented hand-edit of a 64-hex hash. skills/reflect/SKILL.md:57-58 now defers to a "repository's skill-authoring workflow" with draft/test/iterate and description-optimization loops that don't exist in this repo (meta-mode/playbooks/authoring-a-skill.md is a 4-step checklist that defers to the harness).

P3 nits

• Leftover `Task` mentions the cleanup missed: skills/meta-mode/playbooks/opening-a-pr.md:5, skills/reflect/references/tooling-reviewer.md:36, judgment-reviewer.md:23, divergent-reviewer.md:24.
• Tautological assertion: scripts/skill-integrity.test.mjs:91-95 (notEqual(digest, digest(content + "mutation")) is always true once the preceding equality passes).
• Non-strict runs print only canonical skill bodies: N problem(s) with no per-skill detail (check-upstream.mjs:365-368).
• README.md:269 — "Strict mode requires" stranded alone by a missed paragraph re-wrap.
• "active Harness delegation interface" adds a fourth synonym alongside delegation API / delegation tool / documented delegation mechanism.
• Coverage gap (pre-existing): the 3 mattpocock/skills-derived skills (codebase-design, diagnosing-bugs, writing-for-agents) have no digest protection — the blanket "detect missing/changed/truncated bodies" claim holds for 47/50 skills.
• --file as the last argument silently falls back to the default path (scripts/model-config.mjs:11-14, pre-existing).

Recommendation

Mergeable as-is. I'd fast-follow items 1–3: make missing baselines a strict failure, extend the ~\ expansion to run-role.mjs/smoke-harnesses.mjs, and add a mutation test covering the canonical check block.

---

Review method: 4 parallel adversarial subagent reviews (correctness/security, tests, docs, manifest integrity), each with independent recomputation and mutation experiments; findings cross-validated before posting.