plan-handoff

Plan Handoff

Safety: never relay a clean-looking handoff over a dirty tree or a red baseline test run — if the script warns, fix the baseline first. --test-cmd runs as a plain subprocess (no shell), but its last output line is echoed verbatim into the prompt, so don't pass a command whose last line could print a secret.

Claude-Code-specific, flagged not rewritten: the disable-model-invocation: true frontmatter field is a Claude Code directive (only trigger on an explicit user request, never automatically). If the host platform has no equivalent, the skill's description text is the only signal it has for when to trigger this skill.

Emit a copy-paste prompt that lets a fresh session implement a reviewed plan against a pinned baseline. The output is a prompt the user pastes into a new session — this session does not execute the plan.

The plan and the repo rules (CLAUDE.md, AGENTS.md, the plan's own header) already carry most context. This skill adds only what they don't: an exact baseline to verify, and out-of-band friction the planning session hit but never wrote down.

Workflow

1. Identify inputs. The plan file path (required) and any companion design/spec/notes file. Confirm both are committed — the baseline pins a commit, so a dirty plan means an ambiguous handoff.
2. Read the plan so you do not restate it. Then collect, from this session only, items that are NOT already in the plan, spec, CLAUDE.md, AGENTS.md, or another repo skill:
   • Pitfalls / friction the planning session surfaced but did not record.
   • Other context that did not make the plan. If there are none, pass none — the script drops empty sections.
3. Find the test command from the repo's rules (e.g. uv run pytest, mvn test). This is what the new session re-runs to confirm the baseline.
4. Run the script — it captures git HEAD/branch/tree, runs the tests once to confirm green, prints the finished block, and also saves it to a session-namespaced file (see Script contract):