[calvin-archastro]

Review on ArchCode

Problem and intent

firstlanding added archdev plan collab tasks publish --handle <seat.json> --file <preview.json> and a Tasks tab that shows the hosted full task-plan review (task tsk_2rlTGJzb34VrCF1Y9tKDhk). The companion skills still told agents to answer plan.tasks.request with a flat plan.tasks.result draft and to hand the person archdev tasks review start. Nothing told an agent to report progress or publish.

Flow today vs after:

before: plan.tasks.request -> agent sends plan.tasks.result (5 fields) -> page shows a command to copy
after:  plan.tasks.request -> run.status working (run = tasks:<request>)
        -> tasks-skill preview -> plan collab tasks publish -> run.status done -> Tasks tab shows hosted review

What changed

1. plan-collab/SKILL.md: "Break into tasks" rewritten. Capability check, run.status with run:"tasks:<request>" and empty ranges (the page ignores task runs with another id, sender or non-empty ranges), build the preview with the tasks skill, publish, close the run done or stopped. Removes plan.tasks.result.
2. tasks/SKILL.md: new "From a plan collab request" subsection pointing at the above; use tasks publish instead of tasks review start.
3. plugins/*/skills/{plan-collab,tasks}/SKILL.md: regenerated with npm run build in tools/plugin-gen (npm run check passes).

Scope

Skills text only (plus generated copies). No code.

Risk

Low. Behavior change is limited to agents that handle plan.tasks.request. Needs a CLI that has plan collab tasks publish.

User impact

After Break into tasks, the person sees the agent's progress and then the full hosted task review instead of a command to copy.

Testing

• tools/plugin-gen: npm run check PASS (156 files match). npm test: 127 pass, 2 fail, both archdev-nudge.sh ... runs in well under 100 ms timing assertions (claude, codex hooks), unrelated to skill text.
• Real-agent dogfood on production archdev.ai, throwaway plan art_034nRgwtw1GBc8W5phLcO8, real CLI (built from firstlanding main at 160cd8143c, see caveat), real browser:
  1. plan collab start --file plan.md -> page URL; clicked Approve plan, then Break into tasks in the page.
  2. Seat received plan.tasks.request with request = 01a17fb7-ba61-7bac-ad93-22c156f3bb02:tasks:1, revision 0.
  3. plan collab run with run = tasks:01a17fb7-ba61-7bac-ad93-22c156f3bb02:tasks:1, state:"working", empty ranges -> {"sent":true}; page switched to "Drafting tasks" with the steps.
  4. Built a 3-task preview (objective, acceptance, verification, implementation, risk, dependencies) and ran plan collab tasks publish -> revision 1, status:"reviewing", destination Platform Tasks.
  5. run.status done; Tasks tab showed "Task plan review - revision 1, 3 tasks" with full detail. Nothing approved or saved to the backlog.
• Observed, not fixed: after the run, the page's approval check "Every agent run reviewed" lists the tasks: run as not reviewed (PlanCollabApp.tsx approvalChecks does not exclude tasks: runs, unlike the run banners). Follow-up in firstlanding.
• Caveat: plan collab tasks publish is in firstlanding main (ca40b7842c, 2026-10-27) but in no release yet (latest v0.55.6). The dogfood used a locally built binary reporting 0.55.6.

Follow-ups

1. The skills' minimum CLI gate (0.49.6 in the bootstrap scripts) is not bumped: no release contains the verb. Bump after the next release; until then the skill tells the agent to check plan collab tasks publish --help.
2. Known limit (tsk_620dJcXk4yzNKXpqUbYzJx): the hosted saver drops attachments, decisions and deploy_target on approve. Skill text says so.
3. Throwaway plan artifact left in place; no delete verb found.

🤖 Generated with Claude Code