testing: no contract tests for full pipeline on Windows/MSYS; encoding and permission-prompt paths unverified

[Stashub]

Summary

The plugin has no automated contract tests that exercise the full
setup → spec → plan → start → finalize pipeline on any supported
OS, and specifically not on Windows/MSYS. Core commands are
effectively untested against:

1. Bash-on-Windows quirks (path separators, line endings, shebang
   resolution, $HOME vs %USERPROFILE%).
2. PermissionRequest hook behavior on Windows (may differ from macOS).
3. Deferred tool loading via ToolSearch (see related issue on
   AskUserQuestion).
4. Non-ASCII file encoding in generated reports — Cyrillic text in
   SPEC.md / KNOWLEDGE.jsonl / FINAL.md risks mojibake if any
   stage assumes UTF-8 but a tool writes CP1251.
5. Filesystem verification in finalize mode (see related issue on
   ghost-file entries in FINAL.md).
6. Setup-time placeholder substitution and complete vs partial
   status reporting (see related issue on brewcode-review SKILL.md
   placeholders).

Contract tests exercising the pipeline on each supported OS would
have caught several existing issues automatically before 3.5.2
shipped. Today each Windows user independently re-discovers them.

Proposed scope

Minimum viable contract test suite, one run per supported OS matrix
entry:

1. Setup contract

Run /brewcode:setup on fixture projects (empty, small-JS,
small-Java). Assert:

• All expected files are created.
• No residual setup-time placeholders remain in generated files
  (grep -r '{[A-Z_]*}' .claude/skills/ should return 0).
• Setup exit status matches actual completion — complete only if
  every phase executed, otherwise partial.

2. Full pipeline smoke

Run spec → plan → start → finalize on a minimal fixture task.
Assert:

• All phases reach completed in the Task API.
• FINAL.md exists and passes a consistency check:
  • Every file listed in the Artifacts Index exists on disk.
  • Every file on disk under artifacts/ is listed.
  • Test counts in FINAL.md match the actual last test run output.

3. AskUserQuestion deferred-tool fallback

Spawn a subagent without AskUserQuestion pre-loaded in its schema.
Assert the skill either:

• calls ToolSearch to load the schema, or
• completes the phase as blocked with a non-success status.

Never skipped + complete.

4. Encoding round-trip

Generate SPEC.md / KNOWLEDGE.jsonl / FINAL.md with non-ASCII
content (Cyrillic test strings). Read them back and assert
byte-identical round-trip, including on CI Windows runners using
git-bash default encoding.

5. Subagent model pinning

Verify that bc-rules-organizer and other post-processing agents
do not silently inherit 1M context from the parent session on
accounts without /extra-usage (see related issue on
brewcode:rules).

Infrastructure

GitHub Actions matrix:

strategy:
  matrix:
    os: [macos-latest, ubuntu-latest, windows-latest]
    include:
      - os: windows-latest
        shell: 'bash'  # git-bash / MSYS2

Gate releases on this matrix passing across all entries declared
as supported platforms.

Why this matters

The plugin's core value proposition is deterministic multi-phase
pipelines with full audit trails. That value collapses if the
underlying tooling fails silently on platforms it's used on.
Contract tests are the only scalable defense — every bug fix to
individual issues will otherwise leak back over time.

Related

• brewcode:rules silently fails on accounts without /extra-usage
• /brewcode:setup Phase 5 silently skipped
• bc-coordinator finalize: FINAL.md lists files that do not exist
• setup: brewcode-review SKILL.md has ~13 unresolved placeholders
• Windows: pipeline runs despite macOS-only declaration

A single contract-test suite addressing items 1-3 above at once
would be the highest-leverage investment. It would prevent all five
of the above issues from regressing.