Ralpha Loop

Execute a deterministic loop:

1. Implement or patch.
2. Run targeted tests/checks.
3. Send the same reviewer agent the updated diff and evidence.
4. Address findings.
5. Retest.
6. Repeat until reviewer reports no findings.

Loop Contract

• Keep one persistent reviewer agent for the whole task.
• Keep one persistent worker agent only if parallel implementation is needed.
• Reuse agent IDs across iterations; do not respawn unless the agent is dead/unrecoverable.
• Require at least one concrete verification step after every code change.
• Treat reviewer findings as the primary queue; do not add unrelated refactors during the loop.

Setup

Create and track an in-memory registry at loop start:

• reviewer_agent_id
• worker_agent_id (optional)
• iteration (start at 1)
• open_findings (list)
• resolved_findings (list)

Spawn reviewer once with scope, files, and review rubric (bugs/regressions/tests). Keep this reviewer alive until completion.

Iteration Procedure

1) Implement

Apply the smallest change set that addresses current scope or findings.

2) Verify

Run fast, relevant checks first:

• Linters/compilation/static checks for touched files.
• Focused tests for modified behavior.
• Runtime smoke test when applicable.

If a check cannot run, record why and continue to review with that gap explicit.

3) Review (same reviewer)

Send reviewer a delta-focused prompt:

• Current objective.
• Files changed this iteration.
• Verification evidence.
• Previous findings status (fixed/pending).

Ask for severity-ordered findings with file:line and explicit “no findings” when clean.

4) Triage

For each finding:

• Confirm reproducibility or code-level validity.
• Mark as accepted or contested with short rationale.
• Patch all accepted findings before next review round.

5) Retest

Re-run affected checks after patches.

6) Loop Gate

• If reviewer reports findings: increment iteration and repeat.
• If reviewer reports no findings: exit loop.

Agent Coordination Quirks

• Prefer send_input to the existing reviewer agent over spawn_agent.
• Use interrupt=true only to redirect stale reviewer work; otherwise queue normally.
• Do not close reviewer agent between rounds; preserve context continuity.
• If reviewer times out, use wait with longer timeout before deciding to respawn.
• Respawn reviewer only when agent is closed, unresponsive across retries, or context is corrupted.
• When respawning is unavoidable, send a compact handoff:
  • latest objective
  • current diff summary
  • findings ledger
  • verification results

Findings Ledger Format

Track each finding as:

• id: stable label (e.g., R1-F2)
• severity
• location (path:line)
• summary
• status (open, fixed, contested)
• evidence (test/check or reasoning)

Reuse the same IDs when reporting progress to reviewer.

Completion Criteria

Exit only when all are true:

• Reviewer explicitly reports no findings.
• Latest checks for touched behavior pass (or remaining gaps are documented).
• Final summary includes:
  • what changed
  • what was tested
  • reviewer outcome
  • residual risks/gaps