Agent Workflow

Give your coding agent a verification loop.

Veritas turns mutation survivors, replay drift, fuzz seeds, assertion candidates, and confidence scores into concrete instructions an AI agent can act on before it declares a change done.

review-ai verify --changed repair-prompt evolve --dry-run

Paste This

Use veritas as the adversarial verification loop for this change.

Install if needed:
  curl -fsSL https://github.com/Jacobious52/veritas/releases/latest/download/install.sh | sh

Before editing broadly:
  veritas review-ai
  Read .veritas/ai/change_digest.md and .veritas/ai/agent_feedback.md.

After making code or test changes:
  veritas verify --changed --profile ci
  veritas score
  veritas repair-prompt

If veritas reports findings:
  1. Run veritas explain <finding-id>.
  2. Inspect .veritas/assertions/ and .veritas/corpus/.
  3. Prefer focused regression tests or fuzz corpus entries before changing production code.
  4. Rerun veritas verify --changed --profile ci.

Do not ignore warning/error findings without explaining why.

Review Loop

1. Scope

Run veritas review-ai. The agent reads changed files, changed symbols, risk hints, and diff context.

2. Verify

Run veritas verify --changed --profile ci after code or test edits. Stay changed-scope unless the task needs broader proof.

3. Score

Run veritas score to summarize remaining risk from findings, mutation survivors, replay, corpus, properties, and budgets.

4. Repair

Run veritas repair-prompt to get a compact AI repair queue with proof commands and done-when criteria.

5. Promote

Turn useful repros and survivors into owned tests through promote-repro or promote-regression.

6. Evolve

Inspect .veritas/evolution/*_suite.json and apply selected candidates only when the follow-up report improves.

AI-Facing Artifacts

  • .veritas/ai/change_digest.md: changed files, symbols, risk, and diff excerpts.
  • .veritas/ai/agent_feedback.md: concrete next-step instructions.
  • .veritas/symbol_graph/*.json: Tree-sitter symbols, owners, line ranges, and risks.
  • .veritas/assertions/*.json: structured assertion candidates.
  • .veritas/corpus/*.json: persistent repro seed metadata.
  • .veritas/mutations/*_campaign.json: mutant status records.
  • .veritas/differential/*.json: behavior replay manifests and results.
  • .veritas/evolution/*_suite.json: ranked next-generation work queue.
  • .veritas/regressions/*.md: generated assertion guidance.
  • .veritas/budgets/*.json: command budget metadata.

Rules For Agents

  • Keep verification changed-scope unless the user asks for a full workspace run.
  • Treat surviving mutants as missing behavior assertions, not automatic production defects.
  • Prefer a small regression test before editing production code when a repro exposes behavior.
  • Do not widen fuzz time, package caps, coverage scope, or reverse dependency depth without explaining the runtime tradeoff.
  • Generated tests are reviewable artifacts, not automatically trusted source.