--- name: super-oracle description: "Run Claude, Codex gpt-5.6-sol, and configurable OpenCode reviews in parallel. Each reviewer gets one refinement pass before one provider writes the final synthesis. Use for deep code review, difficult cross-file debugging, architecture or refactor planning, research synthesis, and multi-model second opinions." --- # Super Oracle Run the council from the repository you want to review: ```bash bash /scripts/super-oracle.sh \ -C /abs/path/repo -o /abs/path/oracle-round-N.md /abs/path/briefing.md ``` By default, three reviewers work separately and in parallel: - Claude Fable runs at `xhigh`. - Codex `gpt-5.6-sol` runs at `xhigh`, with fast mode and the fast service tier. - OpenCode runs Kimi K3 at `max`. Each healthy reviewer then checks all initial reports and writes one refined report. The latest Opus combines the best available reports into one synthesis. These flags change the run: ```text -s claude|codex|opencode skip a provider, repeat as needed, kimi aliases opencode -y opus|codex|opencode choose the synthesizer, kimi aliases opencode -m fable|opus choose the starting Claude reviewer -q 1..100 set the Claude usage cutoff, default 93 -O MODEL choose the OpenCode model, default Kimi K3 router -V VARIANT choose the OpenCode variant, default max -n disable Claude/Codex MCP for this run -t use cheap test models, never an oracle verdict ``` Test mode has fixed models. `-m`, `-O`, and `-V` are invalid with `-t`. The wrapper skips a missing CLI and prints a warning. A run fails if an explicitly chosen synthesizer is unavailable. When the default Opus synthesizer is unavailable, the wrapper may use a healthy Codex or OpenCode provider. Every run needs `jq` and `pgrep`. Production Claude runs also need `curl`. ## Write the briefing The reviewers do not know the current conversation. Start with `reference/briefing-template.md`. Include: - State the exact goal and question. - Give the working directory and review scope. - Include the specification and repository instructions. - List the relevant files, commands, and current test results. - State the constraints and non-goals. - Record the decisions from earlier review rounds. The wrapper gives each reviewer `reference/code-review-principles.md`. These principles are general engineering guidance. They are not a checklist or a required report format. ## Read the output The `-o` file contains one synthesis for you to judge. The wrapper prints every active path before it starts the models. With the default artifact directory, the paths are: ```text OUTPUT.artifacts/reviews/claude.md OUTPUT.artifacts/reviews/claude.partial.md OUTPUT.artifacts/reviews/claude.fable.partial.md OUTPUT.artifacts/reviews/claude.refined.md OUTPUT.artifacts/reviews/claude.refined.partial.md OUTPUT.artifacts/reviews/claude.refined.fable.partial.md OUTPUT.artifacts/reviews/codex-gpt-5.6-sol.md OUTPUT.artifacts/reviews/codex-gpt-5.6-sol.partial.md OUTPUT.artifacts/reviews/codex-gpt-5.6-sol.refined.md OUTPUT.artifacts/reviews/codex-gpt-5.6-sol.refined.partial.md OUTPUT.artifacts/reviews/opencode.md OUTPUT.artifacts/reviews/opencode.partial.md OUTPUT.artifacts/reviews/opencode.refined.md OUTPUT.artifacts/reviews/opencode.refined.partial.md OUTPUT.artifacts/reviews/synthesis.md OUTPUT.artifacts/reviews/synthesis.partial.md OUTPUT.artifacts/reviews/synthesis.fable.partial.md ``` The three `*.fable.partial.md` files get content only when an interrupted Fable attempt has a checkpoint. The wrapper does not create files for skipped providers. Use a new output file and artifact directory for every review round. When a refinement fails, the wrapper uses the labeled initial report for synthesis. It does not copy the initial report into the refined file. The refined file may keep partial text. OpenCode stops on quota, token, rate-limit, authentication, billing, context, or maximum-token errors. The wrapper keeps its partial text and diagnostics. It does not call OpenCode again in that council. Other providers continue when they have usable reports. A result marked `DEGRADED` has reduced coverage. It cannot prove convergence. A synthesis failure ends the run with an error and keeps the source files. ## Decide what to fix 1. Read the synthesis. Read source reports when you need to check evidence, disagreement, or review coverage. 2. Judge every finding against the repository. Model agreement is not proof, and you may reject a finding. 3. Fix only accepted findings. Make the smallest sufficient change and test it. 4. Run a new production council with the same providers. Add short notes about earlier decisions when a repeated finding could cause churn. 5. Continue until a non-degraded run has no finding that you agree should be addressed. Rejected findings do not keep the cycle open without new evidence. The same is true for duplicate, speculative, cosmetic, or out-of-scope advice. Ask the user when a finding needs a product decision or new authority. ## Keep the run safe Child processes bypass permission checks and sandboxing. Use a trusted working directory. Claude may use only first-party subscription authentication. The wrapper clears Anthropic API credentials and gateways. It checks subscription usage before and during production Claude stages. A paid overage event stops the Claude stage. Reviewers may use subagents for separate investigations that matter to the change. Each reviewer chooses those investigations from the code and the engineering principles. The prompt blocks forced checklist fan-out, unrelated workflow work, and recursive Super Oracle calls. Claude `Agent` and `Workflow` stay enabled. Claude child skills, slash commands, and session persistence stay disabled.