oracle

name

oracle

description

Use when Codex should escalate a difficult problem to ChatGPT in a confirmed Pro extended-thinking mode, or the closest explicitly confirmed high-intelligence fallback, through the @Browser plugin. Trigger when ordinary Codex reasoning has not produced a satisfactory path, when Codex would otherwise hand a hard decision back to the user but a stronger reasoning pass may justify continued autonomous progress, or when a blocker, complex design/research/debugging/synthesis task, ambiguous judgment call, or high-value decision would benefit from a slower one-shot answer. Default to a fresh ChatGPT chat with a self-contained context packet in @Browser, confirm the exact checked model/mode label before sending, submit once, then use a 10-minute heartbeat automation to check for the result instead of blocking or resubmitting. When the answer arrives, summarize it, verify locally, continue the work, and delete the automation. Do not send secrets or sensitive private data without explicit user approval.

Oracle

Purpose

Use Oracle to spend a stronger, slower external reasoning budget on problems that normal Codex work has not resolved well enough. The default interaction is a fresh, one-shot ChatGPT conversation opened through the @Browser plugin: provide a complete context packet, submit it once, register a 10-minute heartbeat automation to check for completion, then convert the answer into verified local progress when it arrives.

Core Orientation

Treat a confirmed ChatGPT Pro extended-thinking mode as an intelligence amplifier, not just a reviewer. Use it for any shape of hard problem where a deeper model may help:

• Diagnose a blocker Codex has failed to resolve.
• Resolve a difficult decision Codex would otherwise ask the user to make.
• Reframe a confused design, research, or architecture problem.
• Stress-test assumptions, issue framing, acceptance criteria, or evidence.
• Synthesize scattered artifacts into a decision.
• Compare paths when the wrong choice would waste significant time.
• Find the missing abstraction, invariant, experiment, proof, or counterexample.

Do not narrow the interaction to debugging only. The useful Oracle question may be diagnostic, adversarial, creative, strategic, mathematical, architectural, empirical, or editorial depending on the task.

Question Quality

Because Pro/extended answers can take a long time, optimize for leverage rather than format. The async heartbeat workflow means long waits are acceptable; do not shrink a high-value question merely to keep the current Codex turn short. Before opening ChatGPT, decide what kind of intelligence you need:

• Breakthrough: "What am I missing?"
• Diagnosis: "Which hypothesis best explains this evidence?"
• Design judgment: "Which approach is safer under these constraints?"
• Adversarial review: "Where is this plan brittle or self-deceiving?"
• Synthesis: "What conclusion follows from these artifacts?"
• Experiment design: "What is the smallest test that separates these explanations?"
• Explanation: "Give me the clearest mental model for this behavior."

Prefer one high-leverage main question when possible, but do not force every request into a single rigid format. If a problem deserves broad context, many artifacts, or careful extended analysis, provide that context and ask for a deep answer with explicit assumptions, uncertainties, falsification checks, and concrete next actions. Do not ask for hidden chain-of-thought; ask for the useful reasoning artifacts needed to continue the work.

If a short exploratory exchange is the right way to reach the real question, use it.

Fresh Chat Default

Start a new ChatGPT chat for Oracle by default. Reusing an old ChatGPT thread usually imports stale context, hidden assumptions, and token pressure that make the answer less reliable.

Use an existing authenticated ChatGPT tab only as a logged-in browser surface. From there, create a new chat or otherwise verify that the composer is in a blank conversation before submitting. Use an existing ChatGPT conversation only when the user explicitly asks for continuity with that thread.

Treat each Oracle call as a self-contained one-shot operation:

• Prepare the context in Codex first.
• Send one complete prompt into a fresh ChatGPT chat.
• Register an async heartbeat automation instead of waiting in a blocking polling loop.
• Check every 10 minutes until one substantive answer is available.
• Extract the useful reasoning back into Codex.
• Verify locally and continue.

This does not forbid follow-ups, but follow-ups are exceptions. Prefer making the first prompt complete enough that no ChatGPT history is needed.

Model And Mode Verification

Verify the selected model and reasoning mode immediately before sending the Oracle prompt. Do not infer that Pro was used from account status, plan text, the word 확장, or a generic high-intelligence label.

Current ChatGPT UI can show the model selector inside the composer near the right side of the input box. The collapsed selector may display a combined label such as Pro 확장 모드; opening it may show separate radio options such as Instant, Thinking • 확장, and Pro • 확장. Treat these as distinct choices:

• Pro • 확장 or an equivalent visible label with both a Pro-tier model and extended/deep reasoning selected is the intended Oracle mode.
• Thinking • 확장 is an extended-thinking mode, but it is not Pro. Do not call it Pro, and do not submit a Pro-required Oracle prompt while this is the checked item unless Pro is unavailable and you intentionally accept it as a fallback.
• The word 확장 only confirms the reasoning budget/mode. It does not confirm the model tier.

Before submitting:

1. Open the model selector if the collapsed composer label is ambiguous, truncated, icon-only, or could be confused with another mode.
2. Identify the checked/selected menu item, not just the hovered item or menu title.
3. Require the checked item or collapsed label to include the Pro-tier name and the extended/deep reasoning indicator for a Pro extended Oracle call.
4. If a non-Pro mode such as Thinking • 확장 is selected, switch to the Pro extended option when visible and available.
5. If Pro extended is not visible or cannot be selected, decide whether the closest visible high-intelligence fallback is good enough for the task. If you use a fallback, record the exact label and report that Pro was not confirmed.

Record the exact visible label in the locator context and later status, for example model selector checked: Pro • 확장 or fallback used: Thinking • 확장; Pro not visible. This is a claim boundary: never say that Oracle used ChatGPT Pro unless the UI actually confirmed a Pro-labeled selected mode before submission.

Before Handing Off To The User

When the next step is difficult mainly because Codex lacks confidence, not because the user must make a preference or authorization decision, use Oracle before asking the user. The purpose is to get enough external reasoning to keep working without an unnecessary handoff.

Good candidates:

• Choosing between two technical approaches when both seem plausible.
• Deciding whether evidence is strong enough to proceed, close an issue, or change direction.
• Selecting the next experiment after local investigation stalls.
• Resolving a reasoning conflict between implementation evidence, docs, and prior assumptions.
• Finding a conservative default when the user has already given a broad goal.

After Oracle responds, continue without asking the user only when the recommendation is actionable, consistent with the user's stated goal, low enough risk, and locally checkable. Record the reasoning briefly in the final answer or durable artifact.

Still ask the user when the decision is about user preference, product/business intent, spending money, sharing or uploading private data, credentials/login/OTP, destructive or hard-to-reverse actions, legal/medical/financial judgment, or anything the user explicitly reserved for themselves.

Context Packet

Build only as much context as ChatGPT needs. Include any of these when helpful, in whatever order fits the task:

• The user's objective and what would count as progress.
• The current blocker or uncertainty.
• Relevant evidence: excerpts, error text, command output, metrics, docs, data slices, or observations.
• What Codex already tried and why it was unsatisfying.
• Constraints, acceptance criteria, risk boundaries, and user preferences.
• Candidate hypotheses or paths, if they exist.
• The type of answer wanted: diagnosis, critique, design choice, next experiment, proof, synthesis, or implementation guidance.

Keep the packet compact, but do not starve the model of decisive context. Redact secrets, private identifiers, credentials, tokens, customer data, and unnecessary full-file dumps.

Async Heartbeat Workflow

Oracle is fire-and-forget by default after the prompt is submitted. The goal is to avoid Codex waiting in a long polling loop and to avoid accidental retries while ChatGPT is still thinking.

After sending the prompt:

1. Record enough locator context to find the pending answer later: ChatGPT tab/conversation URL if visible, chat title if visible, exact checked model/mode label, whether Pro extended was confirmed or a fallback was used, submission time, and a one-line summary of the question.
2. Create a heartbeat automation attached to the current thread that runs every 10 minutes. Use the automation tool when available; prefer a thread heartbeat over a detached cron job.
3. Make the automation prompt self-contained. It should say to open or inspect the existing ChatGPT Oracle conversation, check whether the answer is complete, preserve the recorded model/mode claim boundary, and never resubmit, reload, regenerate, interrupt, or duplicate the prompt while the answer is pending.
4. If the answer is still pending, leave the automation active and keep any user-facing status brief.
5. If the answer is complete, extract the answer, summarize the useful content, distinguish Oracle suggestions from locally verified facts, perform the next local verification/action where feasible, then delete the matching heartbeat automation.
6. If ChatGPT shows a clear failure state rather than a long-running response, report that state and decide locally whether a retry is justified. Do not retry simply because more than 10 minutes elapsed.

Use a distinctive automation name such as Oracle result check: <short task> so the completion run can identify and delete the correct automation. If automation tools are unavailable, fall back to sparse manual polling in the current turn, but still submit the prompt only once.

Because this workflow removes most waiting pressure, be willing to use Oracle for harder, broader, and more synthetic questions than would fit a synchronous Codex turn. The constraint is prompt quality and data safety, not answer latency.

Browser Workflow

Use the @Browser plugin to interact with ChatGPT. If the Browser plugin skill is available, load and follow it before browser actions. In current Codex skill listings this may appear as Browser Use: browser or browser-use:browser; use that plugin-associated skill rather than shell browser automation, web.run, standalone Playwright scripts, macOS open, or Computer Use.

1. Open or select ChatGPT. Prefer an existing authenticated chatgpt.com tab as the browser session; otherwise open https://chatgpt.com/.
2. Start a fresh ChatGPT chat before submitting. Use the visible new-chat control or another visible blank-conversation path, and verify that the prompt will not be appended to an old transcript.
3. If login, password, OTP, account selection, or browser permission is required, ask the user to complete it. Do not handle credentials or bypass checks.
4. Select and verify the visible Pro extended mode before sending. Because UI labels can change, rely on the visible checked/selected model-menu item and the collapsed composer label, not on old instructions. The current UI may place the selector on the right side of the input box and distinguish Thinking • 확장 from Pro • 확장; choose Pro • 확장 for a Pro Oracle call. If only a non-Pro high-intelligence fallback is available, record the exact fallback label and do not claim Pro was used.
5. Send the prepared self-contained question/context once. Do not duplicate the same prompt while a long response is pending.
6. After submission, switch to the async heartbeat workflow. Do not keep the current turn open solely to poll ChatGPT unless the answer is already visibly complete.

Treat the user's explicit request to use Oracle as permission to send a bounded, redacted technical prompt. Still confirm before sending sensitive material or nontrivial private data.

Conversation Strategy

Prefer a single strong answer from the fresh chat. Use follow-ups only when they clearly increase expected value:

• Add a new test result and ask which hypothesis now survives.
• Ask the model to attack its own recommendation.
• Ask for the smallest decisive next step.
• Ask it to choose between two concrete paths under the user's constraints.
• Ask it to restate uncertainty and falsification evidence.

Avoid low-value follow-ups that only ask for more volume. If the answer is vague, make the next question sharper rather than broader.

If a follow-up is needed and may take time, treat it as a new one-shot async Oracle step: submit once, update or recreate the heartbeat with the new pending-question context, and continue checking at the same 10-minute cadence.

Returning To Codex Work

Do not treat ChatGPT's answer as authoritative. Use it as candidate intelligence to verify locally.

When reporting back:

• State what was asked and the exact model/mode label that was confirmed, including whether Pro extended was confirmed or a fallback was used.
• State whether the answer came back through the heartbeat automation or immediate observation.
• Separate Oracle suggestions from locally verified facts.
• Name the next action, decision, or issue/doc update the answer supports.
• Preserve uncertainty when the answer depends on assumptions not yet checked.
• Delete the Oracle heartbeat automation after the answer has been processed.
• If the task has a durable issue, doc, report, or PR context, write the distilled conclusion there instead of leaving it only in chat.