claude-agent-blueprint

Onboards a new or existing project: scans the codebase and generates the skill set + docs. Use when setting up the agent system in a fresh repo, or re-scanning an existing project to refresh its skills and documentation.

Executes a single planned task end-to-end on its own branch: implement, white-box review, black-box test, learn, merge. Use when a task (FEAT/BUG/REFAC) is planned and approved and you want it fully implemented and merged.

Captures post-task learnings and writes actionable insights back into the agent skill files. Use after a task finishes to record what worked or failed so the agents improve over time.

Handles a small, low-risk change directly with a short mandatory plan, bypassing the full task pipeline. Use for minor edits (typo, copy, tiny fix) that don't warrant a full plan/review/test/merge cycle.

Reference overview of the Claude Agent System: architecture, the command/skill pipeline, state management, and quality gates. Use when you need to understand how the agent system fits together; this is reference material, not an executable workflow.

Detects merge conflicts, analyzes their root cause, and creates resolution tasks for the pipeline (it does not resolve conflicts directly). Use when a branch merge hits conflicts that need systematic handling.

White-box code review of an implementation against its plan: checks code quality, conventions, plan-version match, and prohibits acceptance-criteria changes. Use right after implementing a task, in parallel with black-box tests.

Read-only project status report: shows the queue, active task, current pipeline phase, and history from orchestrator_state.json. Use when you want to inspect progress without changing anything.

Plans a task with validation: defines acceptance criteria, edge cases, and a test spec, then sanity-checks the plan against the codebase. Use before implementing any non-trivial feature, bugfix, or refactor.

Black-box acceptance testing derived only from the plan, independent of the implementer: verifies behavior against acceptance criteria and checks plan-version. Use right after implementation, in parallel with white-box review.

Analyzes failing tests one by one and decides per test whether to fix the code or the test, re-running after each code fix. Use when review or test reports test failures.

Deep validation of a task plan against the actual codebase (8-check suite). Use to vet a plan before execution -- manually, or auto-triggered for complex tasks.