[mauricios]

Why

A developer running several sessions comes back to one and can't tell what it's waiting for. The report after the draft pull request was told to stay quiet. Nothing noticed a merge made on the host. Only a merged pull request could close up. And some stops came before the risk did. Decision 0047 keeps the framework's gates and makes every session move toward done, and then to archived.

What changed

• Every turn ends with what's next (AGENTS.md § AI interaction rules). Reports, numbered options, and recommendations stay as they are. The closing says who acts next: the agent carrying on, the developer with the decision or test the work waits on, or something outside. When the work is done, it offers the step that closes it. A person is asked only at a gate.
• Gates relaxed:
  • Careful lane: the yes comes once, before the push, not before the commit. /adf:open-pr asks for it.
  • Change surface: implementation stops only when the change grows into risk (a new layer, a new dependency, or an escalation trigger). A file outside the plan's table is recorded and work goes on.
  • Moving up a lane: moving into careful carries on with its checklist. Moving into full still stops.
• The pull request report names the developer's next step once. Still no reminders, no offers to promote, no reviewers requested.
• New pr-state.sh hook (SessionStart, UserPromptSubmit). It says once when the branch's pull request merged (offer /adf:wrap-up) or closed without merging (ask whether the task is dropped).
  • It calls gh at session start, then at most once every five minutes.
  • It's silent without gh, and PR_STATE in config.sh turns it off.
• /adf:wrap-up closes up every ending:
  • Merged: as before.
  • Answer delivered: no pull request and no commits.
  • Abandoned: the developer confirms the task is dropped. The branch on origin is kept.
  • On hold: the reason and what resumes it are recorded, and the task isn't closed up.

Where: the skeleton's AGENTS.md, specs/README.md, plan template, Cursor rule, claude-code.md, git-workflow.md, settings.json, and config.sh. The adf plugin's hooks and seven skills. Onboarding and the skills reference. The CHANGELOG entry has the upgrade impact.

Verified / not verified

• Verified: ./evals/run-evals.sh, all suites green (static 202, hooks 115, modules 96, plugin 51, upgrade 45). There are nine new pr-state.sh tests against a fake gh, and new static checks for 0047. build-committed.py --check round-trips.
• Not verified:
  • The hook hasn't run in a real Claude Code session against a live pull request.
  • The paid session evals weren't run; they're for release time.

🤖 Generated with Claude Code