Drydock

Plan-first parallel execution for Claude Code: a rigorous plan document as the source of truth, subagents executing it in parallel waves with disjoint file ownership, a conformance audit gating every wave, and a reconcile loop that feeds execution learnings back into your docs.

**Status: open pilot (v0.8.4).** Four plans have been planned, executed and gated with it — see the [case study](docs/case-study-001-homepage.md) and [docs/plans/](docs/plans/). Gate compliance is measured, not asserted: 27 of 28 wave gates invoked at their boundary across those 4 plans, **1 skipped** — and the skip is the useful part, because the next gate refused to open on the missing report and the retroactive audit found a real ownership breach behind it. Every session knew it was observed, so the figure is a ceiling rather than a rate, and it is published as PUBLISHED, **not** PASSED. From v0.6.0 ownership is **enforced by a hook** rather than requested in prose — and as of 2026-08-22 that hook is observed denying a real edit in a live session, not merely unit-tested (A6). A plan can also carry a **Testing Gate**, which `seatrial` drives against the running app through Playwright MCP, capturing the evidence each case declares and closing with a go/no-go sheet; it has run end to end on this site (A7), where its three designed-to-fail cases all failed as designed. What one run costs is [published](docs/cost-001.md) — including the number that run never captured. Behind it: an adversarial [self-audit](docs/self-audit.md) of the auditor itself (it caught a lying executor while all tests were green, and exposed + fixed a defect in our own attribution design) and an honest [compatibility checklist](docs/compatibility.md) of what's verified vs pending.

Homepage: · Plugin: [`drydock/`](drydock/) · Lifecycle, p