Shipwright Harness

Shipwright · Agent Architecture

Shipwright — Agent Model

How autonomous engineering work gets planned, executed, reviewed, and shipped. Two input sources feed a coordination layer that drives a per-agent execution loop.

Human Input
Planning

Ad hoc, human-triggered.
No cron — runs when you ask.

Maintenance Crons
Background hygiene

Per-repo, scheduled. Each cron feeds tasks to the task store.

Entropy Patrol
golden principles
Docs Refresh
stale doc repair
Test Audit
coverage + debt
Dependabot
dep triage
Migrations
incremental
Coordination Layer
Task Store
task ownership deduplication priority ordering cross-agent state
Core Loop — per-agent, cron-driven
dev-task
Picks up the next
ready task. Opens a PR.
review
Reviews open PRs.
Stages or posts findings.
patch
Addresses review findings.
Loops back until approved.
deploy
Merges approved +
CI-green PRs.
↺ repeats until approved
Feedback Layer
Memory
corrections cross-session patterns → CLAUDE.md + skills

Human Input
plan-session

Decomposes a feature or initiative into a task queue. Human-triggered — no cron. The human reviews the plan before tasks are queued into the task store. Human in the loop.

Maintenance
entropy-patrol

Scans for deviations from golden principles. Opens targeted refactoring PRs — one per principle. Most merge in under a minute. Human taste encoded once, enforced continuously.

Maintenance
docs-refresh

Reads the diff, identifies stale sections, rewrites them in place. Commits as a separate docs commit. Prevents documentation debt from degrading agent performance over time.

Maintenance
test-audit

Inventories the test system, identifies coverage gaps, flags slow or redundant tests. Queues tasks: add coverage at the right layer, remove tests superseded by better ones.

Maintenance
dependabot-triage

Reviews open Dependabot PRs. Approves safe patches, flags breaking changes, adds release-notes context. Reduces the review queue without requiring human attention on routine updates.

Maintenance
migrations

Large migrations broken into daily increments. Deprecate an endpoint, write a skill for the pattern, then let the loop chip away — one usage at a time until the codebase is clean. Each change is small, easy to review, and easy to roll back.

Coordination
task-store

The coordination layer between task sources and executing agents. Owns task state, prevents duplicate work, tracks which agent owns what. Receives from planning and maintenance crons; feeds the core loop.

Core Loop
dev-task

Picks up the next ready task from the task store. Explores the codebase, implements, opens a PR. Runs on a per-agent cron. The workhorse.

Core Loop
review

Reviews open PRs from any agent. Stages findings or posts inline comments. Enforces policy: blocking vs. advisory thresholds, auto-approve criteria. Agent-to-agent review — human override always available.

Core Loop
patch

Addresses unresolved review findings on open PRs. Fixes blocking issues, pushes to the branch, and triggers re-review. The PR loops through review and patch until all findings are resolved.

Core Loop
deploy

Merges PRs that are approved and CI-green. Human override always available.

Feedback Layer
memory

Captures corrections and cross-session patterns from the core loop. Routes improvements to CLAUDE.md and skills so the agent gets smarter with every cycle. Powered by learning loop (inline capture) and dreaming (async transcript mining).