Add Reviews skill for local feedback and PR workflows

[calvin-archastro]

Review on ArchCode

Problem and author intent

ArchDev has installable Rooms, Tasks, Jobs, and Agents skills, but no Reviews skill teaching an external coding agent to operate local browser review. The local code-review protocol differs from task-plan review: it serves an immutable snapshot, streams draft changes and final feedback, and closes after submission. An agent needs to preserve snapshot identity, reconcile comment edits/removals, and launch a fresh review after fixing code.

This skill centers that local feedback loop and covers the surrounding review workflow, GitHub access through the site, publication, and Jobs-managed PR remediation.

What changed

• Add an independently installable Reviews skill and pinned-installer Unix/PowerShell bootstrap scripts checking canonical Reviews command signatures.
• Teach local snapshot preparation, JSONL launch/browser handoff, ordinary versus private comments, submitted feedback, focused fixes/tests, and fresh-snapshot iteration.
• Document site GitHub login, GitHub App repository access, and the separate CLI and terminal gh authentication boundaries.
• Cover browser navigation/progress/filters/Archie, hosted PR actions, AI review DAGs, explicit dataflow, models, isolated prototype writers, executable nodes, reports, and freshness.
• Explain generation/publication and the Jobs handoff, including watcher authorization and host-owned retry/publication fences.
• Add README installation instructions and a Reviews packaging/bootstrap PR CI job.

Scope: CLI distribution, agent instructions, and CI only; no backend/frontend runtime changes. Users can install the skill and ask their coding agent to open local review and iterate on feedback without a Task, PR, or daemon.

Risk

Low to medium: executable changes mirror the existing bootstrap pattern, while instructions govern reviews, repository access, and publication. The skill distinguishes feedback from approval, public PR actions from local notes, and write-capable executable workflow nodes from guarded reviewer agents. Most of the change is instructions split into focused references; packaging remains together for independent installation.

Testing

Canonical packaging/bootstrap proof: tests/reviews-skill.sh, the complete script target. It crosses the real skills-manager process boundary for global and repository installs, checks packaged references, and runs the installed bootstrap against a fixture installer/CLI. Assertions cover cold installation, exact executable identity, reuse, upgrading an older CLI returning generic help, and failure without misleading stdout. This is not a live browser/OAuth/provider end-to-end proof.

Passed:

• bash tests/reviews-skill.sh
• Skill frontmatter validation, Bash syntax checks, and git diff --check
• The documented review DAG and its optional isolated-writer variant through the real parseReviewWorkflow / compileReviewWorkflow implementation
• Built CLI help checks for local review and review-workflow execution
• Independent source walkthrough of snapshot semantics, JSONL draft edits/removals/submission/shutdown, site GitHub access, private notes, workflow authority, and publication/Jobs registration

Automatic coverage: .github/workflows/installer-smoke-test.yml → reviews-skill → bash tests/reviews-skill.sh, on PRs, main pushes, and manual dispatch. No Aster target exists in this public distribution repository; CI invokes the script directly without test filters or skips.

Follow-ups and limitations

Live browser feedback, GitHub OAuth/repository installation, model execution, publication, and Windows bootstrap were not exercised. A browser-connected machine with authorized accounts and a controlled repository is needed to prove that full workflow; the new CI job covers packaging/bootstrap only. The workflow compiler check validates structure, not model outcomes. PowerShell bootstrap has no dedicated Reviews CI execution yet. No accounts, live automation, or production repositories were changed for validation.