chore(ship): one PR per unit, auto-chained, gated only on explicit decisions

[BorisTyshkevich]

Summary

• /ship's documented default squashed every remaining phase of a multi-phase issue (or every issue in a comma-list) into one shared PR at the end of the run. In practice this was never actually followed: ADR-0005: adopt @clickhouse/client-web behind the SQL Browser transport adapter #585 (5 phases) and Umbrella: V2 architecture refactor — shell primitives, composition root, state reactivity, transport adapter #593 (8 phases) both carry an explicit owner decision overriding it to one PR per phase, matching how Separate Library queries from dashboard-owned query copies #427 shipped historically. Every time, a human had to notice this and manually re-invoke /ship ISSUE.N phase-by-phase.
• There was also no structured "gate" concept — /ship 585 stopped after Phase 0 partly because of a legitimate review-loop exhaustion, and partly because the issue's own prose ("Phases 2-4 ... do not proceed without a new decision") meant no further phase would ever auto-continue, but nothing in the skill recognized that prose as a first-class stop condition.
• Restructures the per-unit cycle so each unit (a phase or a whole issue) gets its own branch, plan, PR, 3-pass code-review loop, and merge, then auto-chains to the next unit off the freshly-merged origin/main with no re-prompting.
• Adds explicit gate detection: a > **Gate:** blockquote convention under the affected ## Phases row(s) as the primary signal, with a prose-scanning fallback (matched against ADR-0005: adopt @clickhouse/client-web behind the SQL Browser transport adapter #585's and Umbrella: V2 architecture refactor — shell primitives, composition root, state reactivity, transport adapter #593's actual live wording) for issues written before the convention existed. A gate stops only its own dependency spine; every other independent unit/spine still ships in the same run.
• Plan-approval exhaustion (5 passes) and merge-proof exhaustion (3 passes) remain whole-run FULL STOPs, unchanged — these are genuine failure signals, not routine phase boundaries, so they don't get the "continue other spines" treatment gates get.

Scope

Process/documentation only — skills/ship/SKILL.md and three of its references/*.md files. No changes to any .workflow.mjs script or to parse-invocation.mjs: none of the four workflow scripts hardcode a "whole run" assumption (each already takes planFile/prUrl/branch as a plain per-call argument), so this is purely a rewrite of the coordinator's own operating contract.

Test plan

• node --test skills/ship/tests/invocation.test.mjs skills/ship/tests/workflow-contract.test.mjs — 8/8 pass, confirming the literal strings (High risk + Large unit, decompose-and-implement-loop.workflow.mjs, etc.) those tests grep for in SKILL.md/per-issue-cycle.md survived verbatim.
• Hand-verified the new gate-detection trigger phrases against the live text of ADR-0005: adopt @clickhouse/client-web behind the SQL Browser transport adapter #585 ("do not proceed without a new decision") and Umbrella: V2 architecture refactor — shell primitives, composition root, state reactivity, transport adapter #593 ("human merge gate between every phase") — both match exactly.
• Full read-through of the edited SKILL.md for internal consistency: every step-number cross-reference (2.2/2.6/2.7/2.8/step 3) resolves correctly across SKILL.md and references/review-loops.md.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LwFPT465eDJqYcRa8HGNLz