feat(skills): add user-invoked workit-architecture

[BrainerVirus]

What

Adds workit-architecture (/wk-architecture), a second user-invoked skill modeled on workit-retro (#210) and adapted from Matt Pocock's improve-codebase-architecture and codebase-design.

• Deepening (default mode): scope by YAGNI. A direction the user names wins; otherwise the hot spots come from git log churn, workit ledger list friction (failed checks and verdicts, rulings, retro: decisions) and workit test-audit. It reads GLOSSARY.md, CODING_STANDARDS.md and ADRs, and never re-litigates an ADR without new evidence. A read-only subagent explores for cited friction and applies the deletion test. The report is 3-5 ranked candidates. Each one has files and modules, friction with 2 or more cited occurrences, the deeper interface, what gets simpler and testable, and a merge-danger line (one-way or two-way door, plus blast radius). The text list is primary; an HTML report is offered only when the host can render one. The chosen candidate is grilled the way workit-shape does it, then handed to a workit fanout plan (scopes and dependencies) or to a single workit-implement slice. It never refactors without the user's approval.
• Instruction files mode: three escalating passes over AGENTS.md / CLAUDE.md: remove no-ops, progressive disclosure into topic files, then coding rules to CODING_STANDARDS.md and mechanical rules to checks. Each pass is its own commit in one PR. workit knowledge lint is the gate after every commit, and no file is ever scaffolded empty.
• Test sweep mode: workit test-audit <paths> over the suite or a hot spot, with the findings handed to workit-test-audit.
• References: references/vocabulary.md (module, interface, depth, seam, adapter, leverage, locality; the deletion test; dependency categories; replace, don't layer) and references/instruction-files.md.

Wiring

• Manifest, wk-architecture alias, trigger architecture, plus one bootstrap route: - architecture (user-invoked; offer it): workit-architecture.
• disable-model-invocation: true and agents/openai.yaml with allow_implicit_invocation: false.
• Generated Cursor skill copy and /wk-architecture command.
• README skills table.

Tests

• skill-set: the user-invoked list is now [workit-retro, workit-architecture], and each user-invoked route must carry (user-invoked; offer it). The existing tests cover the description, triggers, Codex flag, references, CLI grammar and canonical cross-skill references.
• methods: the skill count is 13, and phrase pins cover the load-bearing doctrine: never refactors without approval, 2 or more cited occurrences, the merge-danger door plus blast radius, each pass its own commit in one PR (SKILL.md and the reference), and never scaffold an empty file.
• Pi alias test: wk-architecture.

Budget

The new description and route line cost about 51 resident tokens. Rebased onto main after #218's bootstrap trim, the resident text measures 1678 of 1700.

Review fixes (be05890)

• Old tests are deleted only after naming the break they catch and showing that a test at the deeper interface fails on that planted break.
• No instruction-files commit adds a workit knowledge lint finding; the lint exits 0 after the last pass, or the findings that remain are reported.
• Each pass runs workit check test and commits through workit git commit, and a sentence that a test pins stays unless the user approves changing the test. The PR is opened through workit-ship. Topic files are checked by hand, since the lint does not read them.
• Churn skips lockfiles, version-bump manifests, generated files, docs and chore(release) commits.
• The explorer is read-only (on Claude Code, the Explore agent): code, git, ledger and test-audit only, never session transcripts or other workspaces, and no secrets copied.
• Each candidate gives an interface sketch, settled during the grilling. Cross-skill references are sentence-final, and a module named for a new concept gets a glossary entry (workit-shape).
• Route phrase: architecture, deepen modules. New pins: never re-litigate an ADR, wait for approval, no commit adds a finding.

Verification on be05890: bun run check (unit 1595 pass, packaging 231 pass, tsc), bun run knip and bun run test:acceptance (24 pass), all green.

🤖 Generated with Claude Code

[github-actions]

🎉 This PR is included in version 7.6.0 🎉

The release is available on:

• v7.6.0
• GitHub release

Your semantic-release bot 📦🚀