fix: align todos-plan with deployment doctrine

[andrei-hasna]

Summary

• replace the retired deployment classification in the canonical todos-plan skill with independent product-story, server-backend, authenticated-authority, and provider-role facts
• preserve the full stable UUID, bounded collection, fresh CAS, protected opaque input, named C1-C9, immutable receipt, compensation, and fail-closed production envelope
• add Codewith to the package-owned agent sync route and allow an explicitly named bundled skill to repair a missing portable-corpus entry

Provenance

The detailed active Codewith copy was manually synchronized on June 30, 2026 (Conversations messages 33579 and 33583; Todos plan 4bc933b6-8b0a-4e38-8807-aa11b034a9af). It had no .hasna-skills.json ownership marker and its detailed content was absent from reachable repository history, so the active copy had diverged without a source-owned update route.

Regression evidence

Red before implementation:

• bun test src/lib/todos-plan-contract.test.ts: 1 pass, 2 fail
• failures proved the canonical skill lacked the required production authority envelope and the named Codewith render route could not resolve bundled-only source

Green candidate:

• python3 .../quick_validate.py skills/todos-plan: Skill is valid!
• focused tests: 54 pass, 0 fail, 206 assertions
• bun run typecheck: pass
• bun run build: pass
• full bun test: 1076 pass, 0 fail, 6628 assertions
• bun run verify:release: pass
• isolated built-CLI Codewith render: rc=0, source and installed SHA-256 both 5513685f72f3c305481d69be0e936dca57f1a76c93f51b092fee918f54a224b5, management marker present
• staged secrets scan: 5 files scanned, 0 findings, not truncated

Scope boundary

This PR does not merge, publish, install, edit profiles, or change the active ~/.codewith/skills/todos-plan copy. Independent review and installed rollout remain with the parent coordinator.

Todos task: 2ba4d4c0-89dd-4525-b4dc-fbc8331e9383

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[andrei-hasna]

[REVIEW] NO_GO — #87 @ 5444e5a — lens: correctness+security+gates, reviewer Trebius (1 of 1)

Reviewed the exact PR base 5520033a662763ec1ce1353363d9e503e76e74d1 through head 5444e5a38876ad91220c3005cdbaac850cd5e81a. I read the full diff for all five changed files and the surrounding sync implementation, portable-skill name normalization and manifest loading, CLI sync argument flow, and existing sync tests.

Commands run, unpiped:

• bun install — exit 0. Setup only; 179 packages installed.
• bun run typecheck — exit 0. TypeScript emitted no diagnostics; this gate has no test pass/fail count.
• bun run test — exit 0. 1075 pass, 1 skip, 0 fail; 6626 expect calls across 1076 tests and 85 files.
• Malformed-name abuse probe against syncSkillsToAgents — exit 0 and printed {"action":{"skill":"../../skills/todos-plan","agent":"codewith","path":"<probe-home>/skills/todos-plan/SKILL.md","action":"create"},"escapedExists":true,"intendedExists":false}.
• Valid-name control for the same bundled fallback — exit 0 and printed {"action":{"skill":"todos-plan","agent":"codewith","path":"<probe-home>/.codewith/skills/todos-plan/SKILL.md","action":"create"},"intendedExists":true,"escapedExists":false}.

Blocking P0/P1 findings:

• P1 — the new explicitly named bundled fallback permits a skill name to escape the declared agent skills directory. normalizeRequested() catches normalization failure and restores the raw name; that raw value then reaches both findBundledSkillSource() and writeManagedAgentSkill() path joins. The reachable input ../../skills/todos-plan resolved the bundled source and wrote SKILL.md plus its ownership marker under the probe home's skills/todos-plan, outside .codewith/skills. The action itself reported the escaped path. With --force, the same path can overwrite an unmanaged SKILL.md outside the target agent directory. This violates the sync command's write scope and is an unsafe mutation path.

Required remedy:

• Reject any name that does not successfully normalize before source lookup or target-path construction. Do not catch normalization errors and reuse the raw input. Add a regression proving a traversal-shaped name performs no write, while the valid named bundled fallback still writes only beneath the selected agent directory.

Non-blocking follow-ups:

• None.

GitHub already reported this PR merged before this review completed. The finding therefore needs a forward fix on the landed code; the green gates do not waive the P1.

[andrei-hasna]

The named P1 is fixed forward in commit 74cce4c23f262860a2ff3027e603477e2a8de3d3 on PR #88.

The repair rejects malformed skill names before any source lookup or target-path construction. The focused regression is now 17 pass, 0 fail; bun run typecheck exits 0; bun run test exits 0 with 1076 pass, 1 skip, 0 fail; the staged security review exits 0 with no findings.

PR #87 was already merged before the review verdict, so #88 is the forward landing path.

[andrei-hasna]

The blocking path-containment finding is fixed and landed through #88.

• Remediation head: 129b9821095a2cd88c47e6731b7e1c2ee0176791
• Independent focused re-review: GO, no P0/P1 findings
• CI: passed
• Squash merge on main: 86c7058a5e65d890a4591bafe52d66207eccdc3f
• Landed tree matches the reviewed merge tree; single Agent: Trebius trailer verified