app-trading-engine: Command for Claude Code

Orchestrate — Automated Review-Fix-Push-Review Convergence Loop

Run the full quality loop until zero changes across all gates: local review, tests, formatting, push, and Gemini review. Produces a session report with every comment's status and industry compliance percentage.

MANDATORY EXECUTION RULES — READ FIRST

These rules are non-negotiable and OVERRIDE any judgment about efficiency or shortcuts:

1. ALWAYS run /review every iteration — spawns 2 fresh agents on ALL changes. Never skip. Never do "quick delta check."
2. ALWAYS run ALL 3 test suites — ./gradlew test, ./gradlew :integration-tests:test, ./gradlew e2e. Never skip even if "nothing changed."
3. NEVER accept or defer ANY comment — dev phase means fix everything. No "accepted," no "out of scope," no "good enough."
4. ALWAYS review ALL changes — every iteration reviews the full diff main...HEAD, not just the delta since last review.
5. ALWAYS wait for and address Gemini review — after push, wait 270s, poll, fix all findings.
6. NEVER terminate early — loop until zero changes in a full pass, or hit max 3 iterations.
7. ALWAYS produce the final report — comment ledger + compliance dashboard + accepted/out-of-scope table (must be empty).
8. ALWAYS write the report to a file — docs/review-reports/{issue}-session-{date}.md for persistence.

Step 0: Initialize

1. Determine the Linear issue number from the branch name or argument (e.g., APP-35).
2. Initialize the comment ledger — an empty list that accumulates every finding across all iterations:

   commentLedger = []
   iteration = 0
   converged = false
   startTime = now
   agentsSpawned = 0
   geminiRounds = 0
   

3. Record the start time for the session report.
4. Record the rollback checkpoint — save the current commit SHA so we can revert if the loop makes things worse:

   BASELINE_SHA=$(git rev-parse HEAD)
   

   If the loop hits max iterations without converging, offer to revert to this SHA.