super-oracle

name	super-oracle
description	Consult Claude Fable in ultracode mode through the Claude Code CLI as an external second-opinion oracle for deep code review, hard cross-file debugging, architecture and refactor planning, and research synthesis. Use for super oracle requests, Claude Fable review, deep review, a rigorous second opinion, or the hard reasoning pass after routine investigation is complete.

Super Oracle (Claude Fable)

Use Claude Fable in ultracode mode as an external second-opinion oracle for the hard 10%: subtle correctness, multi-file debugging, architecture decisions, and research synthesis. Ultracode combines xhigh reasoning with dynamic workflow orchestration. Do routine reads, searches, and edits directly first. Treat the oracle's output as advice and verify load-bearing claims.

Invoke

Write a self-contained briefing, then run the bundled wrapper by absolute path:

bash <this-skill-dir>/scripts/super-oracle.sh -o /abs/path/oracle-out.md /abs/path/briefing.md

Read the answer file and any artifacts it lists before acting. The wrapper uses:

claude --dangerously-skip-permissions -p --effort ultracode --model fable

It also enables the ultracode session setting, verifies Claude Code is at least 2.1.203, disables child skills to prevent self-recursion, and disables session persistence so every call starts fresh. The wrapper requires debug evidence of both a Fable dispatch and ultracode's auto-mode gate; it fails closed if Claude degrades to plain xhigh or workflows are disabled. Set SUPER_ORACLE_REQUIRE_WORKFLOW=1 when a caller specifically needs proof that a dynamic Workflow subagent ran. Pass -n only when configured MCP servers cause trouble; it gives the child a strict empty MCP configuration.

The wrapper is subscription-only: it removes any inherited ANTHROPIC_API_KEY and requires claude auth status to report first-party OAuth before launching Fable. If subscription authentication is missing, run claude auth login; the wrapper will not fall back to Anthropic API credits.

The child Claude process bypasses permission checks. Run the wrapper only in a trusted working directory, constrain the briefing explicitly, and inspect its changes before keeping them.

Hard prompts can take minutes. Progress appears on stderr every 30 seconds, with the first heartbeat within about 15 seconds. Change the cadence with -p SECONDS or SUPER_ORACLE_PROGRESS_INTERVAL; use -p 0 to disable it. Progress never touches the -o answer.

Write the briefing

Assume zero prior conversational context. Include the goal and your role; the absolute working directory; files or PRs to read first; one sharp question; what you already ruled out; constraints and non-goals; and the expected output shape. Demand verbatim evidence and file:line locations for code findings. Tell the oracle whether it may modify files and whether it may fan out. Start from reference/briefing-template.md.

Important contracts

• Fixed model and mode. Keep --model fable and --effort ultracode.
• Workflow permission. Dynamic workflow subagents may fan out for parallel reading and verification. They must not invoke this wrapper recursively.
• Prompt over stdin. The wrapper streams the injected contract plus the briefing over stdin, avoiding shell-quoting and command-length hazards.
• Fresh, non-recursive child. Keep --no-session-persistence, --disable-slash-commands, and the SUPER_ORACLE_ACTIVE recursion guard.
• Success is strict. Require both a zero Claude CLI exit and a non-empty answer. A partial answer is preserved when the CLI fails, but the wrapper exits non-zero.
• Artifacts are deliberate. The oracle may write supporting files only under SUPER_ORACLE_ARTIFACTS_DIR (default OUTPUT.artifacts/). The wrapper appends a manifest to the answer. Read those files before deleting them. Put -o under /tmp or override the artifacts directory to keep them outside the repo.

Consume the output

Verify claims, apply must-fixes, weigh optionals, and reject weak advice. For an approval loop, fix the issue and run a short confirmation briefing. Keep useful briefings and answers as an audit trail.