pigeon

Constraint — read this first

No persona invoked by this skill ever spawns a subagent — see ~/.claude/skills/_shared/pipeline-contract.md "No subagents, anywhere". Every persona switch is this session adopting new operating instructions and writing/reading a file, never a dispatch. This skill reads _shared/pipeline-contract.md for repo resolution, <Name> derivation, the telegraph convention and cost discipline, and _shared/complexity-scoring.md is never consulted directly — every model-tier gate below belongs to whichever skill a persona invokes.

Usage: /pigeon "<task description>" [--name <Name>] /pigeon --resume <Name>

0 — Resolve

1. Resolve the target repo by walking up to the nearest .git, per pipeline-contract.md "Repo resolution". If there is none, say so plainly and stop.
2. Derive <Name> exactly per pipeline-contract.md "Name derivation" (ADO id, or a kebab-case slug of the task description); --name overrides it.
3. Run folder: .claude/pigeon/<Name>/. Create it if absent.
4. Fresh start (no --resume): current persona = task-researcher, input = the raw task description. --resume <Name>: read RUN_LOG.md's last row and the highest-numbered handoff file in the run folder; current persona = that handoff's Next: persona.

1 — Persona registry

2 — The hop loop

Repeat until a handoff's Next: is "none — run complete":

1. Telegraph one line naming the persona about to run (pipeline-contract.md's telegraph convention).
2. Load personas/<NN>-<id>.md for the current persona and adopt its instructions exactly.
3. Give it the latest handoff doc (or the raw task description, on hop 1) plus whatever durable artifact paths that handoff points to.
4. Execute it. It produces its substantive output and writes a new handoff file in the run folder, numbered the next integer in call order (never reusing or overwriting a number), using templates/handoff.md. Append one row to RUN_LOG.md.
5. Read the new handoff's Next: field.
   • "none — run complete" → exit the loop, go to §5 (cleanup).
   • Otherwise, classify the hop: forward if the named persona sits strictly later in the §1 table than the current one, backward otherwise (same or earlier).
6. Hard pause — stop and show the user the handoff before continuing, when any of:
   • the hop is backward;
   • the current persona is business-aligner (its checkpoint always pauses, even forward);
   • the current persona is cybersecurity-expert (always pauses, regardless of direction). On anything but "continue", stop and leave the run folder as-is.
7. Loop-safety (bounce cap): before continuing, count consecutive hops in RUN_LOG.md that landed back on the same persona-id without an intervening forward hop past it. If this exceeds 3, pause and ask the user how to proceed instead of continuing automatically — do not cycle indefinitely.
8. Set current persona = the named Next: persona and go to step 1.

3 — Conditional skip: ui-designer

The skip decision is business-aligner's own (persona 2's file), not a second check here: it reads the plan's File Impact Manifest (from the plan-feature run it just did) and routes Next: straight to software-developer when no touched file matches a recognizable UI path/extension, noting the skip in its Reason: line. pigeon adds no separate detection logic.

4 — --resume

A resumed hop is confirmed exactly like a fresh backward hop (§2 step 6) when the resume target sits at or before where the run paused — never silently resumed forward. The bounce cap (§2 step 7) counts across a resume exactly as it does within one continuous run; RUN_LOG.md is the record, not session memory.

5 — Cleanup on completion

When a handoff's Next: is "none — run complete" (only ever written by ship): telegraph one line that cleanup is starting (not a pause — this is tidying the run's own already-done scratch files, not a forward-progress decision), then delete every numbered handoff file in the run folder, keeping only RUN_LOG.md.

Rules

• Never spawn a subagent, anywhere in this skill.
• Every persona's model tier is inherited from whichever skill it invokes; this skill adds no gate of its own.
• Handoff numbering is call order, never persona order; a file is never overwritten.
• The test engineer's Docker requirement (persona 5) is enforced in that persona's own file, not here — this skill does not pre-check for Docker.