kibitz

name

kibitz

description

Kibitz gets you a second opinion on a plan, sprint plan, spec, or architecture doc using local file-reading CLI agents (Codex and Claude Code) when the Antigravity UI is the driver. The Antigravity lane is the live UI session, never an `agy` CLI subprocess launched from the base system. Use when the user wants a second opinion / to pressure-test / harden / "round-robin" / "make bulletproof" a doc with their LOCAL agents, or says "kibitz", "/kibitz", "kibitz this", "get a second opinion", "run the local panel", or "use Codex, Claude, and Antigravity to review this".

Kibitz Roundtable (Antigravity UI-driven)

When you (Antigravity UI) are the active driver and are asked to run a Kibitz round, you write the grounded anchor review and act as the judge/synthesizer.

Core Rules

1. Active Driver Role: You (Antigravity UI) write the initial grounded anchor review and serve as the final judge/synthesizer.
2. UI-only Antigravity lane: Do NOT run a base-system CLI roundtable for Antigravity. Do NOT spawn agy, --only agy, --all-agents, or any second Antigravity CLI reviewer from this plugin. The live Antigravity UI is the Antigravity panelist.
3. External Reviewers: Codex (codex exec) and Claude Code (claude -p) run as the only external CLI reviewers.
4. No Duplicate Driver: Do NOT duplicate the active Antigravity UI driver through agy CLI. If the user says "all three", treat that as Antigravity UI anchor + Codex + Claude Code.
5. Command Pattern: Run the following command for r1:

   python C:\Users\jeffr\.codex\skills\kibitz\scripts\kibitz.py --doc <plan.md> --round r1 --repo <repo> --driver agy
   

6. Claude CLI spend tier: If the user wants cheaper/faster or deeper Claude review, set KIBITZ_CLAUDE_BUDGET=low|medium|high|plan before the command. Default medium keeps the normal Sonnet/high lane; high opts into Opus/max. This is a Kibitz-controlled tier, not native Claude auto-routing.
7. Round Selection:
   • For r2, r3, and r4, change only the --round value (e.g., --round r2).
   • Run all 4 rounds in order: r1 (high-level) -> r2 (coding plan) -> r3 (wiring) -> r4 (convergence).
8. If the user asks for an Antigravity CLI reviewer:
   • Do not launch it from this plugin. Explain that Antigravity is already present through the UI driver, then continue with Codex + Claude Code unless the user explicitly moves the work to a non-UI Kibitz context.

Step-by-Step Loop

For each round r1 to r4:

1. Write Anchor Review: Read the real source files in the repo. Write your own grounded anchor review of the current plan (following the structure in the corresponding round prompt file under C:\Users\jeffr\.codex\skills\kibitz\references\review-prompt-r<N>.md). Label every claim as CONFIRMED, MISREAD, or UNVERIFIABLE against the actual files.
2. Fan Out: Execute the kibitz.py --driver agy command above to generate reviews from Codex and Claude Code only. Do not invoke agy CLI from the base system.
3. Ground and Verify: Read the generated reviews at <repo>/kibitz-runs/<date>-<topic>/<round>/<agent>.md. Ground and verify every claim made by Codex and Claude Code against the actual files. Discard any hallucinations or misreads.
4. Synthesize: Merge your anchor review with the verified claims from the other agents. Save the resulting updated plan as r<N>_plan.md (or final.md for r4). Keep a short judgment log of accepted, rejected, and verify-at-build items.