oracle

name	oracle
description	Use the @steipete/oracle CLI for debugging, refactors, design checks, and cross-validation, preferring the GCP-backed `oracle-vm` route for agent-driven prompt and literal text/source reviews while preserving explicit host-local Oracle workflows.

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 (VM-first browser, ChatGPT GPT‑5.6 Sol + Pro)

Default agent workflow here: use oracle-vm for prompts plus literal text or source files when research-cpu-01-ts and its remote wrapper are healthy. It runs the supported Oracle wrapper on the GCP VM with ChatGPT GPT‑5.6 Sol and the account/UI Intelligence effort set to Pro. Pro is a separate ChatGPT effort, not a gpt-5.6-pro CLI identifier. This is the human-in-the-loop path: it can take about 10 minutes to 1 hour; expect a stored session you can reattach to or harvest.

Plain oracle remains host-local. Use it for an explicit local request, a VM outage, Mac-only browser context, or inputs that oracle-vm safely refuses. Do not transparently reinterpret a manual oracle command as an SSH request.

Recommended defaults:

• Agent route: oracle-vm -p "<task>" [--file <literal-text-file> ...]; it preflights SSH and the remote wrapper, stages only supported files with protected permissions, creates a slug when needed, and prints same-session reattach and harvest commands. It defaults ORACLE_SOURCE_MAIN=1 for both primary and recovery commands so they prefer the same pinned implementation.
• Engine: browser (--engine browser)
• Model: --model gpt-5.6-sol with --browser-model-strategy select; the signed-in ChatGPT profile supplies the separate Pro effort.
• Browser mode: reuse --remote-chrome 127.0.0.1:9223 when the supported login desktop is already running; otherwise use manual login with --browser-manual-login --browser-chrome-path "$HOME/pro/botfiles/bin/oracle-chrome-linux".
• Local package route: oracle-vm and no-model browser requests through plain oracle prefer the pinned source build at ~/pro/lab/tools/oracle-main commit ea8b1b57f140f2c641a2a8a9cc1dd10bd03bdb18 (upstream PR #320) when it is present and verified. If unavailable, npm latest may be used only with the same requested model and engine. Use ORACLE_SOURCE_MAIN=0 to force npm, ORACLE_SOURCE_MAIN=require to require the pinned build, and ORACLE_SOURCE_MAIN_VERBOSE=1 to print the selected route.
• Verification: confirm the browser/session evidence shows GPT-5.6 Sol and Pro. If the foreground detector misses a visibly completed answer, use oracle-vm session <slug> --harvest for a VM-routed run or oracle session <slug> --harvest for an explicit local run before declaring failure.
• Attachments: directories/globs + excludes; avoid secrets. For text/code-heavy context, prefer compact prompts or inline delivery (--browser-inline-files / --browser-attachments never) before upload mode; reserve uploads/bundles for PDFs, images, binaries, or file sets that truly cannot fit inline.
• Fallback: do not silently downgrade no-model work to GPT-5.5/5.4 or switch it to an API route. Fix or surface browser/profile/canary issues and retry GPT-5.6 Sol + Pro. Honor another model or engine only when the user explicitly requests it.

Golden path (fast + reliable)

1. Pick a tight file set (fewest files that still contain the truth).
2. Preview what you’re about to send (--dry-run + --files-report when needed).
3. Run with oracle-vm when its preflight succeeds so the GCP wrapper selects ChatGPT GPT‑5.6 Sol + Pro by default. Use plain oracle deliberately for the local exceptions above. Use API only when explicitly requested.
4. If the run detaches/timeouts, reattach to or harvest the stored session; don’t start a duplicate run.

ChatGPT Project routing

Oracle can create a new browser conversation inside an existing ChatGPT Project by passing that project's exact /project URL through --chatgpt-url. Treat the project URL as a context scope: the new chat inherits the project's available files and instructions, subject to that project's ChatGPT memory/settings.

Private project names, aliases, and URLs must not be stored in this public skill repository. Before a project-scoped Oracle request, inspect ~/pro/personal_os/context/projects.md for a matching entry under its private Oracle/ChatGPT project-route registry.

Routing rules:

• When the user names a project and exactly one private registry entry matches, automatically add that entry's project URL; do not ask the user to repeat it.
• Do not set a project URL as a shell-wide or wrapper-wide default. Unrelated or ambiguous Oracle work must use the normal non-project route unless the user supplies a URL.
• Do not copy private project identifiers or URLs from the registry into botfiles, logs intended for publication, or public issue/commit text.
• Project placement does not guarantee that prior chats are available as memory. When project context materially affects the answer, ask Oracle to identify whether supporting context came from a project file, project instruction, or prior project chat, and verify distinctive facts rather than accepting a generic claim.
• Project conversations are intentionally not auto-archived by Oracle's default browser archive policy.

Project-scoped GPT-5.6 Sol route on the VM (account/browser default supplies the Pro effort; Oracle does not expose pro as a CLI thinking-time selector):

ORACLE_SOURCE_MAIN=require ORACLE_SOURCE_MAIN_VERBOSE=1 oracle \
  --engine browser \
  --remote-chrome 127.0.0.1:9223 \
  --chatgpt-url '<project-url-from-private-registry>' \
  --model gpt-5.6-sol \
  --browser-model-strategy select \
  --browser-attachments never \
  --slug project-oracle-consult \
  -p '<task>'

Before using that VM route, probe http://127.0.0.1:9223/json/version. If unavailable, start and keep the existing ~/pro/botfiles/bin/oracle-browser-login-vm desktop alive, use its localhost-only noVNC tunnel for login/Cloudflare if needed, and then retry the same --remote-chrome route. Do not silently switch to an API run.

Commands (preferred)

• Show help (once/session):

  • oracle-vm --help
  • oracle --help
  • ORACLE_SOURCE_MAIN_VERBOSE=1 oracle --version

• Preview routing and inputs (no model run):

  • oracle-vm --route-preview -p "<task>" --file path/to/file
  • oracle-vm --dry-run summary -p "<task>" --file path/to/file
  • 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/**"

• VM-first browser run (main agent path; long-running is normal):

  • oracle-vm -p "<task>" --file path/to/file

• Explicit local text/code browser run:

  • oracle -p "<task>" --browser-inline-files --file "src/**"

• Remote session recovery:

  • oracle-vm status --hours 72
  • oracle-vm session <slug> --render
  • oracle-vm session <slug> --harvest

• Explicit default browser run:

  • oracle --engine browser --remote-chrome 127.0.0.1:9223 --model gpt-5.6-sol --browser-model-strategy select -p "<task>" --file "src/**"

• Explicit GPT‑5.4 Pro API override (only when the user requests it):

  • oracle --engine api --model gpt-5.4-pro -p "<task>" --file "src/**"

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

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

• Inspect or require the pinned source route:

  • ORACLE_SOURCE_MAIN=1 ORACLE_SOURCE_MAIN_VERBOSE=1 oracle --version
  • ORACLE_SOURCE_MAIN=require ORACLE_SOURCE_MAIN_VERBOSE=1 oracle --version

Attaching files (--file)

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

That full grammar applies to host-local oracle. oracle-vm v1 intentionally accepts only repeated literal regular text/source files. It rejects directories, globs, exclusions, comma lists, symlinks, common secret-shaped filenames, binary files, files over 1 MiB, and bundles over 8 MiB. Create a minimal sanitized text file or use explicit local oracle when an unsupported input is genuinely required.

• 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/**").
  • Default cap: files > 1 MB are rejected unless you raise ORACLE_MAX_FILE_SIZE_BYTES or maxFileSizeBytes in ~/.oracle/config.json.

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.

Engines (API vs browser)

• The botfiles wrapper on the selected host forces no-model requests onto the GPT-5.6 Sol + Pro browser route. Explicit engine/model requests remain authoritative.
• 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).
  • Use --browser-inline-files / --browser-attachments never for text and source-code context when upload readiness is flaky. If you see Attachments did not finish uploading before timeout, retry with inline files or a compact prompt before changing models.
• 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).
• Browser runs save durable files under ~/.oracle/sessions/<id>/artifacts/, including transcript.md, Deep Research reports, and downloaded ChatGPT-generated images when available.
• Runs may detach or take a long time (browser + GPT‑5.6 Sol + Pro often does). If the CLI times out, do not re-run; reattach or harvest.
  • VM list: oracle-vm status --hours 72
  • VM attach: oracle-vm session <id> --render
  • VM harvest: oracle-vm session <id> --harvest
  • Explicit local equivalents use plain oracle.
• Use --slug "<3-5 words>" to keep session IDs readable.
• oracle-vm creates a timestamped slug when a primary run omits one and prints oracle-vm session <slug> --render|--harvest commands for the same GCP session.
• 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”).

“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 “just enough context”: fewer files + better prompt beats whole-repo dumps.