feat: expert review loop — replace self-review with full adversarial review + fix loop

.github/workflows/agent-fix.md

@@ -42,7 +42,7 @@ safe-outputs:
     max: 3
     target: "*"
   dispatch-workflow:
-    workflows: [polypilot-integration, verify-build]
+    workflows: [polypilot-integration, verify-build, expert-review]
     max: 3
 
 timeout-minutes: 90
@@ -65,9 +65,8 @@ You are an autonomous agent that takes a GitHub issue and delivers a fully teste
 3. **Implement the fix** — make targeted changes
 4. **Run tests** — verify nothing is broken
 5. **Create a PR** — with the fix and test coverage
-6. **Self-review** — do a multi-model adversarial review of your own changes
-7. **Fix review findings** — iterate until clean
-8. **Run integration tests** — dispatch end-to-end verification
+6. **Dispatch expert review + CI** — 3-model adversarial review + cross-platform builds + integration tests
+7. **Post summary** — comment on the issue with results
 
 ## Step 0: Triage the Issue (Go/No-Go Gate)
 
@@ -169,42 +168,20 @@ Then create the PR via `create_pull_request` with:
 - Title: `fix: <description> (fixes #${{ github.event.issue.number || inputs.issue_number }})`
 - Body: description of what was changed and why, linking to the issue
 
-## Step 6: Self-Review
+## Step 6: Dispatch Expert Review and Integration Tests
 
-Launch 3 parallel sub-agents to review your own changes:
-
-```
-task(agent_type: "general-purpose", model: "claude-opus-4.6", mode: "background",
-     description: "Reviewer 1", prompt: "<diff + review instructions>")
-task(agent_type: "general-purpose", model: "claude-sonnet-4.6", mode: "background",
-     description: "Reviewer 2", prompt: "<same>")
-task(agent_type: "general-purpose", model: "gpt-5.3-codex", mode: "background",
-     description: "Reviewer 3", prompt: "<same>")
-```
-
-Each reviewer should check for: regressions, security issues, bugs, race conditions, missing edge cases. Wait for all 3 to complete.
-
-Apply adversarial consensus:
-- 3/3 agree → include finding
-- 2/3 agree → include
-- 1/3 only → discard
-
-## Step 7: Fix Review Findings
-
-For each finding from the self-review:
-1. Make the fix
-2. Run tests again
-3. Commit with descriptive message
-
-Repeat Steps 6-7 up to **3 times** (max 3 review rounds).
-
-## Step 8: Dispatch Integration Tests
-
-After all fixes are committed, dispatch the integration test workflows.
+After all fixes are committed, dispatch the expert code review and integration test workflows.
 
 **Important:** Use the exact branch name from the PR. If you named your branch `fix/issue-N`, the safe-outputs job will use that name without modification (because `preserve-branch-name: true` is set). If you're unsure, use `get_pull_request` to read the PR and get the `headRefName` field.
 
 ```
+dispatch_workflow({
+  "workflow": "expert-review",
+  "inputs": {
+    "pr_number": "<PR number>"
+  }
+})
+
 dispatch_workflow({
   "workflow": "verify-build",
   "inputs": {
@@ -223,21 +200,22 @@ dispatch_workflow({
 })
 ```
 
-## Step 9: Post Summary
+The expert review runs a 3-model adversarial code review (Opus + Sonnet + GPT) on the PR and posts findings as review comments. If it finds issues, the **fix-review-findings** workflow will automatically pick them up, push fixes, and re-dispatch the expert review — looping until zero issues are found.
+
+## Step 7: Post Summary
 
 Post an `add_comment` on issue #${{ github.event.issue.number || inputs.issue_number }} with:
 - What was changed and why
 - Link to the PR
 - Test results (unit tests passed/failed count)
-- Review summary (findings found and fixed)
 - Integration test dispatch status
+- Expert review dispatch status
 - **For visual changes:** note that screenshots are available in the integration test CI artifacts (link to the workflow run)
 
 ## Rules
 
 1. **Fix only the reported issue.** Don't fix unrelated problems.
 2. **Always run tests** before creating the PR.
 3. **Never modify `.github/` files** — protected-files will reject it.
-4. **Max 3 review rounds.** After 3 rounds, post remaining findings as PR comments.
-5. **Never force-push.** Only add commits on top.
-6. **One commit per logical change.** Keep git history clean.
+4. **Never force-push.** Only add commits on top.
+5. **One commit per logical change.** Keep git history clean.