## Summary
- add repository-wide quality tooling and verification scaffolding, including CI workflows, pnpm workspace setup, ESLint/Prettier/markdown checks, and generated-output verification helpers
- reorganize skill sources and generation flow by introducing canonical `_source` variants, generator/manifests, reusable helper abstractions, and shared web-automation/browser utilities
- clean up and expand documentation so the root README flows into docs and skill docs, with clearer development, reviewer, installer, and workflow guidance
## Notable changes
- docs flow and consistency cleanup across `README.md`, `docs/README.md`, and related docs
- new scripts for `check`, docs verification, generated-file verification, shell portability, and safe directory replacement
- refactors in Atlassian and web-automation skill runtimes to reduce duplication and centralize reusable code
- changelog, development documentation, and CI surface updates
## Test Plan
- [ ] `pnpm run check`
- [ ] review generated/manifests and skill sync outputs
- [ ] smoke-check docs flow from `README.md` to `docs/README.md` to skill docs
## Notes
- this branch currently includes tracked `skills/web-automation/shared/node_modules` content that should be reviewed carefully as potentially noisy/accidental committed artifacts
Co-authored-by: Stefano Fiorini <stefano.fiorini@firsthorizon.com>
Reviewed-on: #1
M7 final — adds CLI version checks (claude --version, codex --version)
to the claude-code and codex Prerequisite Check sections so docs and
skill requirements agree. This addresses the only P3 finding from
the M7 review (non-blocking docs-to-skill mismatch).
All 7 milestones delivered across 6 local commits:
- 437b202 M1+M2: claude-code canonical + M1 canonical specs
- d69da3a M3: codex variant
- f404792 M4: cursor variant
- f5161f5 M5: opencode variant
- 9853d49 M6: docs/DO-TASK.md + README updates
- this commit (M7): final gate + P3 fix
do-task ships four runtime variants (claude-code, codex, cursor,
opencode) that execute a single user-supplied prompt end-to-end
with:
- Plan review loop and implementation review loop, each against a
reviewer CLI (codex, claude, cursor, or opencode) with up to
MAX_ROUNDS=10 revisions until APPROVED.
- Parameterized shared Review Loop subroutine (9 steps) invoked
twice per run with loop-distinct session IDs so reviewer context
never leaks across loops.
- Per-payload secret scan (10 anchored regexes; no caching) with
redacted-only user surfacing (`[REDACTED:<label>:<n>-chars]`).
- TDD-first execution via superpowers:test-driven-development with
strict auto-skip limited to pure-documentation and
pure-comment-whitespace-rename.
- Verification gate (lint/typecheck/tests) before the
implementation review loop.
- Single task commit after implementation APPROVED; explicit "yes"
required to push.
- Telegram notification on every terminal outcome.
- Persistent ai_plan/YYYY-MM-DD-<slug>/task-plan.md with 10-value
Status enum and resume semantics.
Reviewer: codex / gpt-5.4. Total review rounds across the 7
milestones: 3+1+1+1+2+1 = 9.
M7 smoke checks (all passing):
- YAML frontmatter: 4/4 parse.
- Core-section schema: 4/4 identical (14 sections, 10 Status enum
values, 15 Runtime State keys).
- docs/DO-TASK.md links resolve.
- 4 variant Phase 1-10 dry-runs coherent; subroutine invoked from
Phases 5 and 8 with distinct SESSION_ID_VAR per loop.
- Opencode reviewer branch passes `bash -n` on rendered Round-1,
Round-N (fresh-call), and Round-N (resume) templates.
- Resume-state walk through Phase 4 reads all 6 session-id keys
correctly.
- Trigger audit: "implement this" only as exclusion; dropped
phrases only under "Dropped defaults".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
M1 canonical specs (opencode reviewer research, task-plan template,
Review Loop subroutine, Phase 1-10 prose, secret-scan regex list)
are embedded in the M2 claude-code SKILL.md as the canonical
reference; later variants (M3 codex, M4 cursor, M5 opencode) will
fork from this file.
Reviewer: codex / gpt-5.4. Approved after 3 rounds:
- Round 1: 2 P1 + 3 P2 → REVISE
- Round 2: 1 P2 → REVISE
- Round 3: 0 P0/P1/P2, 1 P3 (non-blocking) → APPROVED
Key design properties:
- Plan-review payload strips Runtime State and Review History to
prevent reviewer session-ID leakage across rounds.
- Secret-scan step 1a redacts matched text to
[REDACTED:<pattern-label>:<match-length>-chars] before any user
surfacing; never echoes raw match content.
- Brainstorming required for any behavior-changing task; auto-skip
limited to pure-documentation and pure-comment-whitespace-rename.
- Phase 3 reviewer config defaults to codex / gpt-5.4 with
MAX_ROUNDS=10 when user opts for defaults; explicit interactive
default also gpt-5.4 for internal consistency.
- Template Metadata records Branch Name and Worktree Path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
stefano
Stefano Fiorini