AI Agent Workflow Collection

Author a task-specific semi-formal done certificate — a structured verification protocol that a separate validating agent runs to decide whether a task is complete. For each task it restates the definition of done as proof obligations, names the exact evidence to collect and the resolution/regression checks to run, and leaves the status and verdict blank for the validator to discharge. Triggers on "create a done certificate", "add done certificates to the plan", or "generate a completion certificate". Works standalone over any task with a definition of done, and is the companion certificate-author for spec-planner task packages — one certificate per task. To run (discharge) an authored certificate, use validate-done-certificate instead.

Review an implemented change with a semi-formal certificate to decide whether it correctly and completely does what was asked without breaking what it touched. The mandatory correctness gate after a spec-builder task is implemented, run by an agent other than the one that wrote the code. Triggers on "semi-formal review of the diff", "is this implementation correct", or spec-builder dispatching a post-implementation review. Produces a CORRECT / LIKELYCORRECT / CONCERNS / BUGGY verdict. For a done-vs-definition-of-done check use validate-done-certificate; for a standalone patch verification or bug hunt outside a spec-builder build, use reasoning-semiformally.

Create, expand, or change formal design specifications for an app, package, or codebase. Triggers on "create/write a spec", "document the design", "promote to a global spec", or proposing/merging a change spec ("draft a change spec", "RFC for X", "merge the change spec"). Output is a numbered, layered directory of markdown plus a JSON Schema sidecar (repo-wide globals + per-package specs); a change spec is a single document under .specs/changes/. To plan a spec's implementation, use spec-planner; to build that plan, use spec-builder.

Review a canonical or change spec for inconsistencies, divergence, and implementation status, using semi-formal certificate templates. Three modes — (1) a change spec vs the canonical spec it targets (broken references, stale targets, contradictions); (2) a canonical spec vs the code (missing or incorrect implementations, and shipped features the spec never captured); (3) a change spec vs the code (whether its delta has shipped — none/partial/implemented — and which gaps remain). Triggers on "review the spec", "does the implementation match the spec", "find spec divergences", or "has this change spec been implemented". For a single planned task's done-ness use validate-done-certificate instead.

Discharge a task's done certificate against its implementation — collect each obligation's named evidence, run its checks, set each status, trace regressions, and derive a DONE / PARTIAL / NOTDONE verdict by the certificate's rubric. The completeness gate after a spec-builder task is reviewed; when a task has no authored certificate, it derives obligations from the DoD checklist and discharges those the same way. It runs the protocol, it does not author it (that is done-certificates' job). Triggers on "validate the done certificate", "is task NN done", or "check the implementation against its definition of done".