Oracle (CLI) — best use

Oracle bundles your prompt + selected files into one “one-shot” request so another model can answer with real repo context (API or browser automation). Treat outputs as advisory: verify against the codebase + tests.

Main use case (browser, GPT‑5.2 Pro)

Default workflow here: --engine browser with GPT‑5.2 Pro in ChatGPT. This is the “human in the loop” path: it can take ~10 minutes to ~1 hour; expect a stored session you can reattach to.

Recommended defaults:

• Engine: browser (--engine browser)
• Model: GPT‑5.2 Pro (either --model gpt-5.2-pro or a ChatGPT picker label like --model "5.2 Pro")
• Attachments: directories/globs + excludes; avoid secrets.

Golden path (fast + reliable)

1. Bias toward more files to avoid missing context; trim only if token limits demand it.
2. Preview what you’re about to send (--dry-run + --files-report when needed).
3. Run in browser mode for the usual GPT‑5.2 Pro ChatGPT workflow; use API only when you explicitly want it.
4. If the run detaches/timeouts: reattach to the stored session (don’t re-run).

Commands (preferred)

• Show help (once/session):

  • oracle --help

• Preview (no tokens):

  • oracle --dry-run summary -p "<task>" --file "src/**" --file "!**/*.test.*"
  • oracle --dry-run full -p "<task>" --file "src/**"

• Token/cost sanity:

  • oracle --dry-run summary --files-report -p "<task>" --file "src/**"

• Browser run (main path; long-running is normal):

  • oracle --engine browser --model gpt-5.2-pro -p "<task>" --file "src/**"

• If cookie sync is flaky, add --browser-cookie-path "$HOME/Library/Application Support/Google/Chrome/Default/Cookies"

• Manual paste fallback (assemble bundle, copy to clipboard):

  • oracle --render --copy -p "<task>" --file "src/**"
  • Note: --copy is a hidden alias for --copy-markdown.

Attaching files (--file)

--file accepts files, directories, and globs. You can pass it multiple times; entries can be comma-separated.

• Include:

  • --file "src/**" (directory glob)
  • --file src/index.ts (literal file)
  • --file docs --file README.md (literal directory + file)

• Exclude (prefix with !):

  • --file "src/**" --file "!src/**/*.test.ts" --file "!**/*.snap"

• Defaults (important behavior from the implementation):

  • Default-ignored dirs: node_modules, dist, coverage, .git, .turbo, .next, build, tmp (skipped unless you explicitly pass them as literal dirs/files).
  • Honors .gitignore when expanding globs.
  • Does not follow symlinks (glob expansion uses followSymbolicLinks: false).
  • Dotfiles are filtered unless you explicitly opt in with a pattern that includes a dot-segment (e.g. --file ".github/**").
  • Hard cap: files > 1 MB are rejected (split files or narrow the match).

Budget + observability

• Target: keep total input under ~196k tokens.
• Use --files-report (and/or --dry-run json) to spot the token hogs before spending.
• If you need hidden/advanced knobs: oracle --help --verbose.

ChatGPT project (required)

Always use the shared Oracle project in ChatGPT:

--chatgpt-url "https://chatgpt.com/g/g-p-6945c64a0f1081918dd9d6c85d59760b-oracle/project"

Notes:

• This is not the same as the local Oracle session storage folder (~/.oracle/sessions).
• Use --browser-url if your CLI expects that flag name.

Engines (API vs browser)

• Auto-pick: uses api when OPENAI_API_KEY is set, otherwise browser.
• Browser engine supports GPT + Gemini only; use --engine api for Claude/Grok/Codex or multi-model runs.
• API runs require explicit user consent before starting because they incur usage costs.
• Browser attachments:
  • --browser-attachments auto|never|always (auto pastes inline up to ~60k chars then uploads).
• Cookie/keychain access (macOS):
  • Browser runs read Chrome cookies via Keychain. If you see "No ChatGPT cookies were applied", ensure ChatGPT is signed in and approve Keychain access for node.
  • If the Keychain prompt disappears too quickly, rerun the command and approve; you can also use --browser-cookie-path to point directly at Chrome's cookie DB.
• Remote browser host (signed-in machine runs automation):
  • Host: oracle serve --host 0.0.0.0 --port 9473 --token <secret>
  • Client: oracle --engine browser --remote-host <host:port> --remote-token <secret> -p "<task>" --file "src/**"

Sessions + slugs (don’t lose work)

• Stored under ~/.oracle/sessions (override with ORACLE_HOME_DIR).
• Runs may detach or take a long time (browser + GPT‑5.2 Pro often does). If the CLI times out: don’t re-run; reattach.
  • List: oracle status --hours 72
  • Attach: oracle session <id> --render
• Use --slug "<3-5 words>" to keep session IDs readable.
• Duplicate prompt guard exists; use --force only when you truly want a fresh run.

Prompt template (high signal)

Oracle starts with zero project knowledge. Assume the model cannot infer your stack, build tooling, conventions, or “obvious” paths. Include:

• Project briefing (stack + build/test commands + platform constraints).
• “Where things live” (key directories, entrypoints, config files, dependency boundaries).
• Exact question + what you tried + the error text (verbatim).
• Constraints (“don’t change X”, “must keep public API”, “perf budget”, etc).
• Desired output (“return patch plan + tests”, “list risky assumptions”, “give 3 options with tradeoffs”).

Start the prompt with a unique run marker (timestamp + short description) so it’s easy to find later if the run detaches.

“Exhaustive prompt” pattern (for later restoration)

When you know this will be a long investigation, write a prompt that can stand alone later:

• Top: 6–30 sentence project briefing + current goal.
• Middle: concrete repro steps + exact errors + what you already tried.
• Bottom: attach all context files needed so a fresh model can fully understand (entrypoints, configs, key modules, docs).

If you need to reproduce the same context later, re-run with the same prompt + --file … set (Oracle runs are one-shot; the model doesn’t remember prior runs).

Safety

• Don’t attach secrets by default (.env, key files, auth tokens). Redact aggressively; share only what’s required.
• Prefer more context over too little; prune only when token limits demand it.