pi-oracle

pi-oracle lets a pi agent send hard, long-running work to ChatGPT.com through the web app, with repo archives, background execution, saved results, and a best-effort wake-up back into pi when the answer is ready.

Status: experimental public beta. Validated primarily on macOS with Google Chrome/Chromium and pi 0.65.0+. Normal oracle jobs run in an isolated browser profile, not your active browser window.

What a successful run looks like

You: /oracle Review the pending changes. Include the whole repo unless a narrower archive is clearly better.

pi-oracle:
  1. preflights local session/auth readiness
  2. builds a context-rich `.tar.zst` repo archive
  3. starts an isolated ChatGPT web runtime in the background
  4. uploads the archive and prompt to ChatGPT.com
  5. saves the response/artifacts under /tmp/oracle-<job-id>/
  6. sends a best-effort wake-up back to the matching pi session

Later: /oracle-read <job-id>

What you are seeing: the local pi agent keeps control of context selection and safety checks, while ChatGPT web handles the expensive second-opinion work asynchronously. If the wake-up is missed, the result still lives on disk and can be read by job id.

Who this is for

Use pi-oracle if you use pi and want a larger asynchronous reviewer, planner, or analyst that can use your real ChatGPT web account instead of an API model.

It is most useful for:

• maintainers reviewing broad repo changes before shipping
• agents that need a slower second opinion without blocking the main pi turn
• migration, architecture, or failure-mode analysis that benefits from a large archive
• follow-up questions that should continue the same ChatGPT thread later

Do not use it for:

• short local coding tasks that pi can handle directly
• projects that must never be uploaded to ChatGPT.com
• non-macOS environments or machines without the required local browser/tooling

Problem it solves

A normal coding-agent turn is the wrong shape for some work: the task may need a large repo snapshot, a slower web model, a real ChatGPT subscription, artifact downloads, or a durable result that arrives after the active turn ends.

pi-oracle makes that workflow explicit and recoverable instead of asking the main agent to manually drive ChatGPT in your real browser window.

What it does

Problem	Capability	Proof in this repo
Hard tasks need more context than a quick turn should gather.	/oracle prompts the agent to preflight, choose a context-rich archive, and submit it to ChatGPT web.	prompts/oracle.md, oracle_submit, archive tests in scripts/oracle-sanity-*
Browser automation should not steal focus or mutate your active profile.	Jobs clone an authenticated seed profile into per-job isolated runtime profiles.	docs/ORACLE_DESIGN.md, extensions/oracle/lib/runtime.ts
Long jobs need durability.	Job state, responses, logs, and artifacts persist under ${PI_ORACLE_JOBS_DIR:-/tmp}/oracle-<job-id>/.	extensions/oracle/lib/jobs.ts, /oracle-read, /oracle-status
ChatGPT auth can expire or drift.	/oracle-auth refreshes the isolated auth seed from a configured local Chromium profile, with recovery guidance.	extensions/oracle/lib/auth.ts, docs/ORACLE_RECOVERY_DRILL.md
Agents need a simple API, not UI-driving instructions.	The package exposes agent-facing tools: oracle_preflight, oracle_submit, oracle_read, oracle_auth, and oracle_cancel.	extensions/oracle/lib/tools.ts

Fastest way to see it work

1. Install

From npm:

pi install npm:pi-oracle

Or from GitHub:

pi install https://github.com/fitchmultz/pi-oracle

2. Check requirements

You need:

• macOS
• Node.js 22 or newer
• pi 0.65.0 or newer
• Google Chrome or another Chromium-family browser
• ChatGPT already signed in to the configured local browser profile
• agent-browser, tar, and zstd available on the machine
• a normal persisted pi session, not pi --no-session

3. Sync ChatGPT auth once

/oracle-auth

This reads ChatGPT cookies from your configured local browser profile and writes an isolated oracle seed profile. It should not automate your active browser window for normal jobs.

4. Submit a tiny job

/oracle Read README.md and package.json. Tell me in five bullets what this package does and who should not use it.

Expected result:

• The /oracle prompt now runs an early oracle preflight before expensive repo reading or archive creation.
• The agent chooses a context-rich relevant archive up to the 250 MB ceiling, not the smallest possible one-file slice when nearby context helps.
• oracle_submit creates or queues a job.
• If local packing is too large, the prompt treats that as a retryable archive-selection failure and narrows automatically before surfacing the problem.
• The job uploads a repo archive to ChatGPT.com, capped at 250 MiB after default exclusions/pruning.
• The response is saved under /tmp/oracle-<job-id>/response.md by default.
• The matching pi session gets one best-effort wake-up when the job finishes.

If the wake-up does not arrive, run:

/oracle-status
/oracle-read <job-id>

Example requests

/oracle Review the current pending changes. Include the whole repo unless a narrower archive is clearly better. Give me a prioritized code review with concrete fixes.

/oracle Read the codebase and explain the highest-risk auth/session failure modes, including what to test before shipping.

/oracle Explain the README guidance for /oracle-clean retention grace. Archive README.md plus any nearby docs or implementation files that help answer accurately.

/oracle-followup <job-id> Tighten the migration plan around rollback risk, and include the most relevant surrounding files/docs as long as the archive stays comfortably within the 250 MiB limit.

How it works

flowchart LR
    A["/oracle request"] --> B["Agent preflights, then gathers a context-rich relevant repo slice"]
    B --> C["Agent chooses context-rich archive inputs"]
    C --> D["oracle_submit builds .tar.zst archive"]
    D --> E["Detached worker clones isolated auth seed profile"]
    E --> F["ChatGPT web receives archive + prompt"]
    F --> G["Response/artifacts saved under oracle job dir"]
    G --> H["Best-effort wake-up to matching pi session"]

Loading

Key design choices:

• Prompt templates own context gathering. /oracle and /oracle-followup tell the agent how to preflight, gather context, choose archive inputs, and then stop after dispatch.
• Tools own execution. oracle_submit builds the archive, admits or queues the job, starts the worker, and returns immediately.
• Auth uses a seed profile. /oracle-auth imports cookies into an isolated seed profile; each job clones that seed into its own temporary runtime profile.
• Follow-ups preserve ChatGPT thread state. /oracle-followup <job-id> ... resolves the prior job's saved ChatGPT URL and submits the next prompt with followUpJobId.
• Wake-up is best effort, storage is durable. A missed wake-up does not lose the result.

Commands and tools

User-facing commands:

• /oracle <request> — prepare context and dispatch a ChatGPT web oracle job
• /oracle-followup <job-id> <request> — continue an earlier oracle job in the same ChatGPT thread
• /oracle-auth — sync ChatGPT cookies into the isolated oracle auth seed profile
• /oracle-read [job-id] — inspect job status and saved response preview
• /oracle-status [job-id] — inspect a job or list recent job ids when no explicit id is given
• /oracle-cancel <job-id> — cancel a queued or active job
• /oracle-clean <job-id|all> — remove temp files for terminal jobs; recently woken terminal jobs may stay retained briefly, and a blocked cleanup returns the next eligible cleanup time

Agent-facing tools:

• oracle_preflight
• oracle_auth
• oracle_submit
• oracle_read
• oracle_cancel

Configuration

Most users can start with defaults. Set an agent-level config only when you need a non-default preset or browser profile.

~/.pi/agent/extensions/oracle.json

{
  "defaults": {
    "preset": "<preset id from ORACLE_SUBMIT_PRESETS>"
  },
  "auth": {
    "chromeProfile": "Default"
  }
}

Notes:

• defaults.preset is the default ChatGPT model preset for oracle jobs.
• The canonical preset ids live in extensions/oracle/lib/config.ts.
• If the packaged default is fine, omit defaults.preset.
• When an agent is unsure which oracle preset fits, it should omit preset and use the configured default model instead of asking by default.
• You usually do not need browser paths unless auto-detection fails.

Custom Chromium cookie sources

Use this only for a Chromium-family browser that the default cookie importer cannot read.

Before running /oracle-auth with this path:

1. Log into ChatGPT in the target browser profile.
2. Fully quit the browser so its Cookies database is stable.
3. Find the profile Cookies SQLite DB path.
4. Find the browser's macOS Keychain safe-storage item account and service name.
5. Configure all of browser.executablePath, auth.chromeCookiePath, and auth.chromiumKeychain in ~/.pi/agent/extensions/oracle.json.

Example Helium config:

{
  "browser": {
    "executablePath": "/Applications/Helium.app/Contents/MacOS/Helium"
  },
  "auth": {
    "chromeProfile": "Default",
    "chromeCookiePath": "/Users/you/Library/Application Support/net.imput.helium/Default/Cookies",
    "chromiumKeychain": {
      "account": "Helium",
      "services": ["Helium Storage Key"],
      "label": "Helium Storage Key"
    }
  }
}

auth.chromeCookiePath must be paired with auth.chromiumKeychain; partial config is rejected so oracle does not silently fall back to another browser source.

Available presets

Preset id	Description
pro_standard	Pro - Standard
pro_extended	Pro - Extended
thinking_light	Thinking - Light
thinking_standard	Thinking - Standard
thinking_extended	Thinking - Extended
thinking_heavy	Thinking - Heavy
instant	Instant
instant_auto_switch	Instant - Auto-switch to Thinking Enabled

oracle_submit accepts canonical preset ids or a matching human-readable preset label. Keep config values on canonical ids.

Outputs and cleanup

• Jobs persist response text, metadata, logs, and artifacts under ${PI_ORACLE_JOBS_DIR:-/tmp}/oracle-<job-id>/ by default.
• Jobs can queue automatically when runtime capacity is full.
• Completion delivery into pi is one-time best-effort wake-up based.
• /oracle-read [job-id] and oracle_read({ jobId }) inspect saved output later.
• /oracle-clean removes terminal job temp files, but can briefly refuse cleanup after a wake-up so the follow-up turn can still read the saved paths.

Privacy and local data

This extension is local-first, but it handles sensitive local and project data:

• /oracle-auth reads ChatGPT cookies from the configured local browser profile.
• oracle_submit uploads selected project archives to ChatGPT.com.
• Responses, logs, and artifacts are written to the configured oracle jobs directory.

Review the code and design docs before using it with private or regulated material.

Current limits

• Experimental public beta, validated primarily on macOS.
• ChatGPT UI, auth, model controls, and artifact download behavior can drift.
• Archive uploads are capped at 250 MiB after default exclusions and automatic whole-repo pruning.
• A real ChatGPT web session is required.
• The README currently uses command-level proof and design docs; no public screenshot or demo GIF is checked into the repo.
• Production hardening should keep focusing on UI drift detection, auth recovery, artifact capture, and environment diagnostics.

Troubleshooting

/oracle-auth fails or says login is required

• Make sure ChatGPT works in the same local browser profile you configured.
• For custom Chromium cookie sources, confirm auth.chromeCookiePath points at that profile's Cookies DB and auth.chromiumKeychain.services names the browser's safe-storage Keychain service.
• Re-run /oracle-auth.
• Agent callers can use oracle_auth({}) once before retrying a stale-auth oracle submission.
• If ChatGPT is half-logged-in or challenge flow state looks weird, finish the login/challenge in the headed auth browser and retry.

Custom Chromium auth says cookies synced but the session is rejected

This usually means the cookie import worked but the source cookies are not the active ChatGPT session you expected.

1. Open the configured browser profile.
2. Confirm ChatGPT works there without logging in again.
3. Quit the browser fully so its Cookies DB is stable.
4. Confirm auth.chromeCookiePath points at that exact profile's Cookies DB.
5. Confirm auth.chromiumKeychain.services names the browser's safe-storage Keychain service for that DB.
6. Re-run /oracle-auth.

You hit a challenge or verification page

• Solve it in the auth/bootstrap browser if prompted.
• Re-run /oracle-auth before submitting jobs again.

You see "Oracle requires a persisted pi session"

• Do not run oracle from pi --no-session.
• Start a normal persisted pi session, then use /oracle again.

A job finished but no wake-up arrived

• Use /oracle-read [job-id] to inspect the saved response preview.
• Use /oracle-status if you need help finding a recent job id.
• Agent callers can use oracle_read({ jobId }).
• Results are still saved on disk even if the reminder turn does not land.

/oracle-clean refuses a terminal job right after completion

• This can happen during the short post-send retention grace window after a wake-up was sent.
• The command returns a Retry after ... timestamp when that guard is active.
• Wait until that time, then rerun /oracle-clean <job-id|all>.

agent-browser, tar, or zstd is missing

Install the missing local dependency and rerun the command.

Auto-detection picked the wrong browser profile

• Set auth.chromeProfile in ~/.pi/agent/extensions/oracle.json.
• For custom Chromium cookie sources, set auth.chromeCookiePath to the exact profile Cookies DB and pair it with auth.chromiumKeychain.
• Re-run /oracle-auth.

You want more details about a failed run

Inspect the job directory under ${PI_ORACLE_JOBS_DIR:-/tmp}/oracle-<job-id>/. The worker log and captured diagnostics are stored there.

Verification

Useful local checks:

npm run check:oracle-extension
npm run typecheck
npm run typecheck:worker-helpers
npm run sanity:oracle
npm run pack:check
npm test
npm run verify:oracle

npm publish is guarded by prepublishOnly, which runs npm run verify:oracle.

For end-to-end local-extension smoke testing, use docs/ORACLE_ISOLATED_PI_VALIDATION.md. That workflow launches isolated pi sessions against this checkout and uses instant or thinking_light, as required by the project validation policy.

Project map

Path	Purpose
extensions/oracle/index.ts	Extension entrypoint
extensions/oracle/lib/	Commands, tools, config, jobs, queueing, runtime, poller
extensions/oracle/worker/	Detached ChatGPT web worker and UI/auth helpers
extensions/oracle/shared/	Shared process, state, job, and observability helpers
prompts/oracle.md	/oracle prompt-template workflow
prompts/oracle-followup.md	/oracle-followup prompt-template workflow
scripts/oracle-sanity-*	Local sanity and archive-safety checks
docs/ORACLE_DESIGN.md	Architecture, lifecycle, queueing, persistence, recovery behavior
docs/ORACLE_ISOLATED_PI_VALIDATION.md	Repeatable isolated pi validation workflow
docs/ORACLE_RECOVERY_DRILL.md	Safe expired-auth recovery drill

Next action

Install the package, run /oracle-auth, then submit the tiny README/package review job above. If you are evaluating the design before running it, start with docs/ORACLE_DESIGN.md.

License

MIT. See LICENSE.