--- name: frontier-llm-consult description: Human-gated frontier-model research router for AutoOPT Stage 2 symbolic fitting, candidate formula/proof discovery, proof planning, and independent model review. Use when Codex must choose between an external chatgpt-pro-session consultation and an explicitly invoked native solve-with-highest-reasoning campaign, preview external outgoing context or approve native repository scope, delegate to the selected route skill, archive durable evidence, and treat model outputs as candidate mathematical objects rather than proofs. --- # Frontier LLM Consult ## Core Rule Route difficult research work through one of two execution routes without weakening its approval gate: - `chatgpt-pro-session`, the sole external consultation route and the default; - `solve-with-highest-reasoning`, a native same-host Codex campaign. The native route is not an external consultation, an independent second-model opinion, or a source of provider diversity. This skill chooses and coordinates the route; it does not replace or duplicate any route skill. Load and follow the selected route skill before execution. For an external route, never send, upload, or submit consultation context before the user approves the route and outgoing context. For the native route, never initialize or run a campaign unless the user explicitly invokes `$solve-with-highest-reasoning` and approves the native route gate below. Treat every model-produced formula, identity, closed form, proof skeleton, blueprint, proof, construction, counterexample, or algorithm as a candidate until its required independent verification succeeds. ## Inputs For AutoOPT Stage 2, start from the Stage 1 artifacts: - BnB-PEP derivation Markdown; - numerical method parameters, stepsizes, or recurrence data; - dual variables, certificate structure, or solver summaries when available; - structural constraints and desired analytical form; - the requested output type: formula, identity, proof skeleton, or proof-planning blueprint. For standalone routed work, collect the same information at the appropriate level of detail: task, context files, constraints, desired output, evidence standard, and what should be archived. ## Route Selection Recommend `chatgpt-pro-session` for external consultation: - the user explicitly asks for ChatGPT Pro, a persistent session, or reuse of an existing ChatGPT conversation; - the consultation is likely to require multiple turns, corrections, or follow-up symbolic fitting; - the context is broad, reusable, or expensive to re-upload; - the task is ordinary or underspecified AutoOPT Stage 2 symbolic fitting; - the user wants a bounded one-shot external review. In that case, start a new ChatGPT Pro session, submit one approved turn, archive it, and do not create a separate one-shot backend. Select `solve-with-highest-reasoning` only when all of these conditions hold: - the user explicitly invokes `$solve-with-highest-reasoning` in the current workflow; - the target is one exceptionally difficult, sharply defined research problem; - direct repository exploration or installed local tools materially benefit the work; - the user accepts active Goal mode, an equivalent persistent mechanism, or the skill's disclosed checkpointed fallback when neither is available, plus its user-confirmed minimum-duration contract and a 180-minute ceiling for each code or solver execution. The duration defaults to eight elapsed wall-clock hours and 28,800 logged active-work seconds; a user override of `h` hours requires both `h` elapsed wall-clock hours and `h * 3600` logged active-work seconds. Before Goal setup, repository grounding, campaign initialization, or clock start, load the native skill and let it resolve and freeze the duration. If the explicit invocation already contains a valid duration, acknowledge the effective wall-clock and active-work floors without asking again. Otherwise, present the native skill's default-duration notice, ask whether to use the default or a custom duration in hours, and wait for an explicit answer. The router must preserve the selected floor and must not weaken or reinterpret it. Explicit invocation makes the native route eligible; it does not waive the native route approval gate. Exclude the native route when the requested purpose requires provider diversity, an independent second-model review, or a bounded consultation. Do not infer explicit invocation from a router recommendation or from approval of another route. In a routed workflow, this gate is an additional AutoOPT authorization boundary; it does not change the native skill's standalone explicit-invocation contract. Default to recommending `chatgpt-pro-session` for AutoOPT Stage 2 when the user's preference is ambiguous and the native route has not been explicitly invoked. This is only a recommendation before the approval gate; it is not permission to upload. Requests for another provider, API or multi-provider routing, or provider diversity are unsupported by this router. Stop and report the unmet requirement rather than substituting either route. ChatGPT Pro can provide an external second-model review; the native route cannot satisfy provider diversity or independence from the current Codex host. Assess availability and eligibility route by route. A resolvable native skill is not eligible without explicit invocation and acceptance of its full campaign contract. If the selected native route lacks Goal mode, an equivalent persistent mechanism, or the permitted checkpointed fallback, reliable capability resolution, or a required local tool, stop and report the exact blocker. Never silently downgrade its model, reasoning effort, or route. If ChatGPT Pro is selected but unavailable, stop without automatically switching to the native route. A user may choose the native route only through a separate explicit invocation and full native gate. If the selected route is unusable, report Stage 2 as blocked rather than improvising a replacement. ## Context Preparation For `chatgpt-pro-session`, build the smallest outgoing package that carries the truth: - a self-contained task prompt; - a short project or mathematical briefing; - exact files or artifacts to include; - desired output markers or artifact format when a structured response is needed; - explicit constraints, including "candidate, not proof"; - archive destination, usually `ResearchLog/auto-opt-runs/-/stage2/` for AutoOPT. Do not attach secrets, credentials, browser state, API keys, private tokens, or irrelevant generated output. Prefer fewer files plus a precise prompt over a large dump. For `solve-with-highest-reasoning`, create no upload package. Pass the exact Stage 2 target, repository root, Stage 1 seed artifacts, admissible output, and approved discovery boundaries to the native skill together with the confirmed duration source, minimum wall-clock hours, and minimum active-work seconds. Let that skill inspect the repository and build its local problem contract and source manifest. It may identify and read additional files within the approved repository and source scope without another route gate. Require renewed approval before changing the target, using another repository root, widening the source boundary, or making any external transfer. Never change or weaken the frozen duration. ## External Route Approval Gate Before a ChatGPT Pro send/upload action, present: ```text Consultation skill: frontier-llm-consult Recommended route: chatgpt-pro-session Route class: external consultation backend Recommendation rationale: Provider/model or session mode: Outgoing files or context manifest: Preview evidence: Secrets check: Archive destination: Exact ask: approve this consultation route and outgoing context ``` Proceed only after explicit approval. If the actual backend manifest later differs materially from the approved file list or context summary, pause and ask for approval again. This outer gate remains controlling when `chatgpt-pro-session` is delegated through Frontier or AutoOPT, even if a direct, explicitly invoked session workflow would otherwise treat invocation as standing upload consent. ## Native Route Approval Gate Before initializing or running a native campaign, present: ```text Routing skill: frontier-llm-consult Selected route: solve-with-highest-reasoning Route class: native same-host Codex campaign, not an independent second-model review Recommendation rationale: Exact target and admissible output: Explicit invocation: Persistence: Capability: Downgrade policy: no silent downgrade Duration confirmation: Duration contract: Execution ceiling: 180 minutes per code or solver execution Repository root and Stage 1 seeds: Discovery boundary: Immutable or excluded paths: Local write scope: Local tools: External-model transfer: none authorized Public-search boundary: ordinary background and standard named theorems only, per the native skill Authoritative campaign archive: ResearchLog/highest-reasoning-runs/-/ AutoOPT Stage 2 archive: ResearchLog/auto-opt-runs/-/stage2/ Exact ask: approve this native route, target, scope, tools, persistence, and duration contract ``` Resolve and name the strongest model and highest supported reasoning setting when the host exposes them. If they cannot be determined, stop before campaign initialization and obtain the user selection or required capability instead of assuming or downgrading. Approval of the native gate authorizes only the approved local campaign scope. It does not authorize external model transfer, destructive actions, purchases, or material scope expansion. ## Delegation For the `chatgpt-pro-session` route: 1. Load the `chatgpt-pro-session` skill. 2. Use directed context when the Stage 1 artifacts are known; use broad context only when a narrow selection would omit necessary information. 3. For bounded one-shot work, start a new session, use one approved turn, and archive that turn as the external consultation evidence. 4. If continuing an existing session, check whether the uploaded context is still fresh enough for the requested follow-up. 5. Archive or reference the session metadata, manifest, prompt, imported response, and turn artifacts under the Stage 2 evidence folder. For the `solve-with-highest-reasoning` route: 1. Confirm that the user explicitly invoked `$solve-with-highest-reasoning`; route recommendation or generic approval is insufficient. 2. Load the `solve-with-highest-reasoning` skill and follow its persistence, capability, repository-grounding, campaign-record, source-boundary, computation, audit, and stopping rules. 3. Before Goal setup or repository grounding, let it resolve and freeze the user-confirmed duration. Preserve that selected floor throughout routing. 4. Hand it the approved target, repository root, Stage 1 seed artifacts, discovery boundaries, admissible output, duration contract, tools, write scope, and archive destinations. 5. Permit local repository exploration and use of installed tools only within those approved boundaries. 6. Do not let the native campaign recursively invoke `frontier-llm-consult`, ChatGPT Web, or another external-model backend. Any such use requires a separately authorized workflow under its own approval and outgoing-context rules. Each route skill is authoritative for its own mechanics. Do not copy its command syntax, browser-control rules, session-state rules, campaign protocol, or failure handling into this skill. ## Native Route Evidence and Status Keep the authoritative native campaign under: ```text ResearchLog/highest-reasoning-runs/-/ ``` For AutoOPT, store these compact handoff artifacts under the run's `stage2/` folder: - `route-approval.md`, recording the approved target, route configuration, discovery and write boundaries, tools, duration contract, and both archive destinations; - `native-campaign-reference.md`, recording repository-relative campaign and evidence paths, campaign status, model and reasoning evidence, SHA-256 hashes, record-validation status, and audit summary; - `candidate.md`, recording exact candidates, assumptions, provenance, evidence status, and any proposed Stage 3 theorem or certificate target. Use repository-relative paths and SHA-256 hashes. Do not copy or symlink the complete campaign into `stage2/`. Map native campaign states into the AutoOPT Stage 2 gate as follows: - `prepared`, `running`, or `paused`: Stage 2 remains incomplete and resumable; - `complete`: archived candidates may proceed to the Stage 2 researcher gate; - `incomplete`: archive the exact-gap report, and keep partial candidates gated; - `user_stopped` or `environment_blocked`: record Stage 2 as interrupted or blocked, respectively. ## Return Check After an external consultation returns: - save the durable transcript or imported response before summarizing it; - extract the candidate formula, identity, closed form, proof skeleton, or blueprint; - compare the candidate against the available numerical and structural data at a basic sanity-check level; - label the result as `Conjectured` or `Needs check` unless independent proof or Lean verification has already been completed; - record failures as useful evidence when a candidate does not match the data. After a native campaign reaches a terminal state, preserve its authoritative record first, then write or refresh the compact Stage 2 artifacts above before summarizing it. Check each candidate against the available numerical and structural data and present the applicable Stage 2 researcher gate. Do not promote a claim because the selected duration floor was satisfied, subagents agreed, audits were recorded, or the campaign record validator passed. A native-route candidate remains `Conjectured` or `Needs check` unless an independently accepted ordinary proof or formal verification supports promotion. Do not ask the model to re-check itself and then report that as verification. Model re-examination is only another candidate-generating step.