Skip to main content
recursive-mode supports delegating audits and code reviews to subagents. When configured correctly, a subagent receives a complete context bundle, returns grounded findings and a verdict, and the main agent verifies that result before accepting it. Delegated review is optional. If subagents are unavailable, the main agent performs the same audit as a self-audit. Audit rigor is not optional — only delegation is.

​
When to use delegated review

  • Phase 3.5 code review — the primary use case
  • Any audited phase where an independent pass would improve quality
  • High-risk or large-scale implementations where a second opinion on requirement coverage matters

​
How it works

1

Generate a review bundle

Use the recursive-review-bundle script to package everything the reviewer needs into a canonical, reproducible bundle:
python scripts/recursive-review-bundle.py \
  --repo-root . \
  --run-id "<run-id>" \
  --phase "03.5 Code Review" \
  --role code-reviewer \
  --artifact-path "/.recursive/run/<run-id>/03.5-code-review.md" \
  --upstream-artifact "/.recursive/run/<run-id>/00-requirements.md" \
  --upstream-artifact "/.recursive/run/<run-id>/02-to-be-plan.md" \
  --audit-question "Which R# remain incomplete?" \
  --required-output "Findings ordered by severity"
The bundle is saved under /.recursive/run/<run-id>/evidence/review-bundles/. The bundle auto-discovers relevant addenda — do not omit them from the delegated context.
2

Pass the bundle to a subagent

Dispatch the subagent with the bundle path and the phase-specific instructions. Valid subagent roles include:
  • phase-auditor — independent pass over draft, upstream artifacts, diff, and requirement coverage
  • code-reviewer — Phase 3.5; checks requirements vs implementation, plan vs implementation, and code quality
  • traceability-auditor — verifies every in-scope R# is explicitly addressed
  • test-reviewer — verifies test adequacy, exact commands, and evidence capture
  • memory-auditor — Phase 8; verifies memory status transitions and router updates
Vague delegation (“review this phase”) is not valid. The subagent must receive a complete context bundle.
3

Subagent returns findings and a verdict

The subagent must return structured findings and one of:
  • Audit: PASS — the phase is ready to lock
  • Audit: FAIL — repairs are needed before the phase can proceed
A result that contains only generic praise with no grounded findings, or that lacks an explicit verdict, must be rejected.
4

Main agent verifies claims against actual files

Before accepting a delegated result, the main agent must verify the subagent’s claims against:
  • the actual worktree diff
  • the actual changed files
  • the actual phase artifacts and review bundle
If the verification reveals issues, record the concrete repair performed. Do not silently accept stale delegated context.
5

Record a subagent action record

Every meaningful subagent contribution must produce a durable action record under:
/.recursive/run/<run-id>/subagents/
The action record captures inputs provided, claimed actions, claimed file and artifact impact, findings, and the acceptance decision. The Phase 3.5 artifact must also record Review Bundle Path in ## Review Metadata.
Never accept a delegated result without verifying the subagent’s claims against actual files, the actual diff, and the actual recursive artifacts. A subagent’s word is not sufficient — the main agent is always responsible for verification before lock.

​
Self-audit fallback

When subagents are unavailable, perform the same audit locally and record it in the phase artifact: 
Audit Execution Mode: self-audit
Subagent Availability: unavailable
Then complete the full audit loop — draft, re-read upstream artifacts, reconcile against the diff basis, record gaps, repair, re-audit, and set Audit: PASS only when the phase is genuinely ready. Self-audit must meet the same standard as delegated audit.

​
What makes a valid review bundle

A bundle is only valid when it includes all of the following:
  • Phase name and artifact path
  • Artifact content hash
  • Reviewer role
  • Upstream artifacts to re-read
  • Relevant addenda
  • Relevant prior recursive evidence
  • Normalized diff basis from 00-worktree.md
  • Changed file list
  • Targeted code references
  • Phase-specific audit questions
  • Required output shape (findings format and verdict)
If any item is missing, do not delegate — run the audit yourself.