workstream

Update durable repo memory after planning, merge, rejection, blocker, or repeated agent mistake.

Convert an approved plan into PR-sized chunk contracts with scope, acceptance criteria, reviewer requirements, and human review focus.

Review a diff for architecture drift, wrong abstractions, boundary violations, coupling, and chunk-scope violations.

Review CI, lint, typecheck, test, coverage, workflow, and package-script changes for weakening or bypass behavior.

Fast-fail or split unreviewable agent work based on scope, size, file types, boundary count, missing evidence, or unclear intent.

Perform read-only repository discovery for an initiative. Use to map existing behavior, files, tests, risks, and unknowns before planning.

Review whether a code change requires documentation updates, examples, migration notes, policy docs, or developer onboarding updates.

Check deterministic proof before expensive reviewer fanout: tests, lint, typecheck, scope, CI integrity, test delta, dependencies, and PR size.

Triage and respond to external reviewer comments such as CodeRabbit, GitHub checks, or human PR review comments.

Plan a large or ambiguous engineering initiative before implementation. Produces intent, discovery, plan, chunk map, and chunk contracts. Use before writing code for non-trivial work.

Capture human intent, rationale, non-goals, alternatives, boundaries, and proof strategy before implementation.

Review an implementation plan before code is written, especially for L1 infrastructure, architecture, auth, payment, or policy work.

Create a human-readable PR trust bundle containing intent, design, scope, tests, CI integrity, reviewer results, risks, and human review focus.

Review Workstream changes for project manager, worker, reviewer, operator, payment, revision, and audit workflow correctness.

Review a diff for correctness, acceptance criteria coverage, edge cases, regressions, and missing tests.

Review a diff for duplicated helpers, missed existing abstractions, redundant logic, and code reuse blindness.

Classify work by blast radius, risk, SLA, model budget, reviewers, and human gates before implementation.

Review a diff for security, auth, permission, payment, data, secrets, prompt injection, and audit risks.

Review a diff like a senior engineer for maintainability, simplicity, readability, operational risk, and long-term ownership.

Execute exactly one approved chunk using bounded implementation, evidence gates, zero-trust reviewers, repair loop, and PR trust bundle. Do not start the next chunk.

Review changed tests for weakened assertions, removed coverage, skipped tests, and tests rewritten to match broken behavior.