oracle

name	oracle
description	Browser-mode-only Oracle consults: bundle a prompt plus selected files and ask a signed-in ChatGPT GPT-5.5 Pro browser session for a second opinion. Use when stuck, debugging hard bugs, reviewing an architecture plan, or cross-checking a substantive diff with large file context. Never use Oracle API mode from Harness Kit. Trigger: /oracle, /consult.
argument-hint	[prompt] [--file <glob>...]

/oracle

Use Oracle only as a browser-backed consult surface. It bundles the prompt and selected files, opens a signed-in ChatGPT browser session, and stores the answer under ~/.oracle/sessions. The lead agent still owns synthesis, verification, and final judgment.

Hard Boundary

• Always pass --engine browser.
• Prefer --model gpt-5.5-pro unless the operator names another browser model.
• Never run --engine api, --provider openai, --no-azure, --models, or API preflight from this skill.
• If an Oracle command would need an API key or per-token billing, do not use Oracle for that work from Harness Kit; use the normal roster/research paths instead.

Workflow

1. Pick the tightest file set that contains the truth. Exclude tests, snapshots, generated output, and fixtures unless they are the oracle.

2. Preview first:

   npx -y @steipete/oracle --engine browser --model gpt-5.5-pro \
     --dry-run summary --files-report \
     -p "<task>" --file "src/**" --file "!**/*.test.*"
   

3. Run only after the preview is sensible:

   npx -y @steipete/oracle --engine browser --model gpt-5.5-pro \
     --slug "<3-5-words>" \
     -p "<task>" --file "src/**" --file "!**/*.test.*"
   

4. If the run detaches or times out, do not re-run. Inspect the stored session:

   npx -y @steipete/oracle status --hours 72
   npx -y @steipete/oracle session <id> --render
   

Good Prompts

Oracle starts cold. Include the repo goal, relevant commands, exact failure, constraints, and desired output shape. Critics get the artifact and the oracle only — never the author's reasoning trail (shared AGENTS.md: Fresh context beats self-review).

Good output requests:

• Return blockers only, with file/line evidence.
• Challenge this plan; list missing proof, wrong assumptions, and safer alternatives.
• Find the likely root cause and the smallest verification loop.

Safety

• Do not attach secrets, .env, key files, tokens, browser profiles, or private customer data.
• Use --dry-run summary --files-report before every real run.
• Treat Oracle output as advisory evidence. Verify with the repo's live oracle, tests, gate, browser route, or command before claiming done.
• Browser mode is cheaper than API billing but can be slow and subject to ChatGPT subscription limits.