Pipeline and quality gates

Spec-driven development for AI coding agents: design first, plan in phases, build with TDD, and release only with evidence.

  1. The four stages
  2. Pick a tier
  3. Quality gates
  4. Review without the pipeline
  5. Audit a whole codebase
  6. See it run

The RDF pipeline: spec, plan, build, ship, with quality gates scaled to scope

The four stages

Stage Command What happens Output
Design /r-spec Discover the problem, research options, write the spec, run a challenge review docs/specs/*.md
Plan /r-plan Read the spec and code, decompose into TDD phases, run a challenge review docs/plans/*-plan.md
Build /r-build [N\|--parallel] The dispatcher runs each phase: engineer implements, QA verifies, reviewer audits Committed code
Ship /r-ship Preflight, verification, release prep, publish, report Pull request

Enter at any point. Have a spec already? Start with /r-plan. Have a plan? Start with /r-build. Each command tells you the next step.

/r-vpe is an optional orchestrator that runs the whole pipeline from one outcome-oriented request.

Pick a tier

Not every change deserves the full process. Choose a tier at /r-spec or /r-plan. RDF suggests one from the request, you confirm it, and pressing Enter keeps --full.

Flag Use it for Spec Plan review Build gates
--full (default) Multi-component work or open design questions docs/specs/*.md Challenge review Set by scope (below)
--quick Well-understood changes to about 3 files, with no open architecture questions Condensed Single review Gate 3 capped at sentinel-lite
--bugfix A defect you can reproduce with a test None, the failing test comes first Schema check only Gate 1 plus a regression-only sentinel-lite

A five-line fix is just /r-plan --bugfix, then /r-build.

Tiers only remove ceremony, never safety. Security-sensitive changes keep QA and the full 3-pass review whatever the tier. Details: canonical/reference/tiers.md.

Quality gates

The dispatcher derives verification depth from what each phase touches, so you never tag a phase by hand.

Gate Who runs it What it checks
1 Engineer Self-report with evidence
2 QA (rdf-qa) The project checks in governance/verification.md: lint, tests, anti-patterns
3 Reviewer sentinel (rdf-reviewer) Adversarial review. Lite is 2 passes (anti-slop, regression); full adds security
4 UAT (rdf-uat) Real-world scenarios, install flows and CLI use from an end-user persona
Scope The phase touches Gates
scope:docs Documentation, changelog or comments only 1
scope:focused One file, or only config and scaffolding 1, 2
scope:multi-file Two or more source files (the default when scope is unclear) 1, 2, 3 lite
scope:cross-cutting Install scripts, CLI entry points, cross-OS logic, breaking changes 1, 2, 3 full
scope:sensitive Security-critical paths, shared libraries, data migration, credentials 1, 2, 3 full

Gate 4 is added at any scope when a phase changes CLI entry points, help text or man pages. Files named in your governance/constraints.md and governance/anti-patterns.md raise the scope, so your own rules steer the gates.

A --full plan with 3 or more phases also ends with a full 3-pass sentinel review of its cumulative diff.

Review without the pipeline

Every gate is available on demand:

Command Use it to
/r-verify Run a QA pass on a diff or scope
/r-review --challenge <file> Challenge a spec or plan before implementation
/r-review --sentinel Audit finished work (2 or 3 passes)
/r-review-answer Route findings to fix, rebut or defer, with an audit trail
/r-test Run user acceptance tests from an end-user persona

Audit a whole codebase

Command What it does Output
/r-audit Three reviewers and QA across all domains, in parallel AUDIT.md
/r-audit-slop Finds dead code and orphan helpers, then a sentinel peer-reviews findings to remove false positives Validated findings

See it run

The demo walkthrough follows a real production change through every stage, with the actual gate results.