[calvin-archastro]

Review on ArchCode

Problem and intent

Plan co-editing (epic "Live plan co-editing") runs a local agent seat, archdev plan collab start, that joins a plan Artifact next to a person in the browser and relays page signals (tab completion, ask, agent request, run undo, file conflict, approval). Nothing tells Claude Code or other agents how to drive that seat: which verb answers which signal, that positions are Unicode code points, that revisions go stale, or that review suggestions are CriticMarkup text rather than channel messages.

Task: Claude Code can drive the local seat through a bundled plan-collab skill. The installer downloads skills from this public archive (bundling was rejected), so the skill lives here.

What changed

1. plan-collab/SKILL.md: the skill. It covers:
   • archdev plan collab verbs: start, signals (bare form waits and acks; --input sends), read, suggest, complete, apply, run.
   • Every signal kind: complete.request, complete.cancel, complete.suggestion, ask.request, agent.request, agent.reply, run.status, run.undo, run.reviewed, file.conflict, file.resolve, plan.approved, plan.reopened, each with a receive, command, response mapping.
   • Code-point positions, revision freshness, CriticMarkup suggestions versus signal-only ghosts, inverse undo, conflict handling, approval and reopen.
2. catalog/archdev.json: registers plan-collab.
3. plugins/{antigravity,claude,codex,copilot,cursor,grok,opencode,pi}/skills/plan-collab/SKILL.md: generated by tools/plugin-gen (npm run build).
4. RELEASE-server-only-risk.md: lists plan-collab/ as a canonical source.

Verbs and behavior were taken from services/rust/archdev/src/plan_collab/{command,daemon,markup,file_sync}.rs in firstlanding and checked against archdev plan collab <verb> --help (0.54.7). The skill includes the shared "Activity and risk ownership" block that generate.test.ts requires of every distributed skill.

Scope: docs-only skill, no code. Risk: low. User impact: agents in all eight harnesses gain the skill after release.

Testing

In tools/plugin-gen:

• npm run typecheck: clean.
• npm run build then npm run check: 148 generated files match the catalog.
• npm test: 125 pass, 0 fail. This includes generate.test.ts "every plugin carries every skill byte for byte, with its mode" and "every distributed skill teaches server ownership while preserving plan/task risk"; both iterate the catalog, so they cover the new skill with no per-skill test. The latter failed on the first draft until the ownership block was added.
• Automatic coverage: .github/workflows/plugins.yml runs npm run typecheck, npm test and npm run check on PRs.
• Not exercised: no live run of the skill against a real seat (see follow-ups).

Follow-ups

1. firstlanding: register plan-collab in the installer name list (services/rust/archdev/src/installed_skills.rs) once this is released. The installer downloads from this archive; do not restore bundling.
2. Real-seat proof on the task: a real agent session must use this skill to answer signals against a running seat (existing real-seat coverage is services/platform/test/archastro/platform_web/channels/artifact_sdk_end_to_end_test.exs, which does not prove the skill). Task tsk_76zFa5B1WPS4VVrtrhsqDu stays open for it.
3. The skill states no minimum ArchDev version; add one once the first release containing plan collab is known.

🤖 Generated with Claude Code