oracle

Oracle operates as a one-shot context injector: it reads your prompt, expands your file globs, filters out noise (node_modules, dist, dotfiles, files over 1 MB), and assembles everything into a single payload.Pick your files — use --file with paths, directories, or globs to include exactly what the model needs.Preview before spending tokens — run --dry-run summary --files-report to see what would be sent and how large the payload is.Choose your engine — --engine browser drives ChatGPT or Gemini via browser automation; --engine api hits the OpenAI API directly (also supports Claude, Grok, Codex).Run and optionally reattach — long browser runs (up to ~1 hour) are stored as sessions under ~/.oracle/sessions; if the CLI times out you can reattach without re-running and re-spending tokens.

File bundling with glob support — --file accepts paths, directories, globs, and negation patterns. Pass it multiple times or comma-separate entries.Dry-run preview — --dry-run summary|full plus --files-report shows exactly what would be sent and estimates token cost before any real request.Browser & API engines — browser mode automates ChatGPT and Gemini; API mode targets OpenAI-compatible endpoints including Claude, Grok, and Codex.Persistent sessions with reattach — long-running or detached runs are stored locally; list with oracle status and reattach with oracle session .Readable session slugs — use --slug to give sessions human-readable IDs instead of random hashes.Manual paste fallback — --render --copy builds the full markdown payload and copies it to the clipboard for manual pasting into any chat interface.Remote browser host — serve a browser instance on one machine and connect the CLI from another using --remote-host and --remote-token.Duplicate prompt guard — prevents accidental re-runs of identical prompts; override with --force when a fresh run is intentional.

LLM API Key (optional) — Required only when using --engine api. When OPENAI_API_KEY is set, Oracle auto-selects API mode. Without it, Oracle defaults to browser mode. Supports: OpenAI API-compatible endpoints.Browser engine (optional, no key needed) — Browser mode drives an existing ChatGPT or Gemini session via automation — no API key required. Supports: ChatGPT (GPT-4, GPT-5.2 Pro), Google Gemini.

Deep bug diagnosis — Attach your entry points, relevant modules, and the exact error output so the model sees the full picture before answering.Refactor planning — Send a tight file set covering the affected surface area and ask for a patch plan with tradeoffs — no hand-waving about files it can't see.Architecture review — Bundle config files, key directories, and a project briefing to get grounded recommendations without the model inventing your stack.Resumable long investigations — Write an "exhaustive prompt" with a full project briefing and file set so you (or a teammate) can re-run the same query days later and get consistent, reproducible results.