advisor

Advisor

Routing Boundary

advisor pulls JUDGMENT in from a different model. It is the inbound half of the cross-runtime pair; handover is the outbound half (it sends mechanical build work OUT to Codex). Keep them distinct:

• Second opinion on a judgment / design fork / decision → advisor (default).
• Adversarial critique of a PREPARED plan → advisor --panel.
• Review a code diff → review.
• Hand unfinished mechanical build work to Codex to DO → handover.
• Sharpen a fuzzy idea by self-interview → grill.

advisor never writes code, never commits, never touches git. It asks another model and returns the outside view; the caller decides. Never auto-incorporate a finding.

Step 0: Locate the seat (zero-config)

Detect the current runtime only to guarantee that the outside opinion comes from a different model family:

• CLAUDECODE=1 in env → Claude seat; reach out through the verified Codex transport.
• otherwise → Codex seat; reach out through the verified Claude transport.

This seat check chooses direction only. Authentication, budgets, and general model-routing policy remain host concerns.

Step 1: Select roles, then probe (fail loud, never silently self-review)

Select the complete role set before probing: default uses the one seat-derived role from Step 2. A Claude-seat panel uses advisor.panel.openai.fast plus advisor.openai; a Codex-seat panel uses advisor.panel.fast plus advisor.panel.deep. Every panel critic therefore differs from the current seat family. For every unique transport in that set, assert the target binary is on PATH. Then read each row's minimum CLI version from lib/model-anchors.md, parse <name> --version, and stop with ADVISOR: <name> <have> is below verified minimum <need> — upgrade required when it is missing, unparseable, or older. If a binary is missing, print ADVISOR: no cross-model binary on PATH (<name>) — judgment stays local, NOT self-reviewed and stop. A self-issued second opinion is not a second opinion; do not fake decorrelation by asking the same model.

If the anchored call itself fails, emit the same banner with the command's actionable error and keep the judgment local. Do not retry through the current model or through an unpinned default.

Step 2: Default — one cross-model consult

Read lib/model-anchors.md, select advisor.openai from a Claude seat or advisor.anthropic from a Codex seat, and pass both its model and effort explicitly. Pass the question + the minimum context + the specific fork. Narrow in, structured out.

• Claude seat → direct codex exec with the advisor.openai anchor and its read-only sandbox guard.
• Codex seat → direct claude -p with the advisor.anthropic anchor, tools disabled, and session persistence off. Keep the call tight: one question, smallest sufficient context, a structured answer back.

Done when the outside view is returned as outside-voice, unincorporated. The caller owns the decision.

Step 3: --panel — blind cross-model critique

Only for a plan that is expensive if wrong (irreversible, outward-facing, multi-day). A daily plan uses the review pass built into sprint, not a panel. The panel is a forcing function, not a ritual.

Spawn two mutually-blind critics, each a different model and a distinct lens (correctness / user outcome / simplicity / missing constraints / failure modes + reversal). Cross-model is the point: same-model critics decorrelate only the prompt, not the model's priors. Pick one likely-failure lens per critic.

• Select the seat-filtered two-role composition from Step 1. Use every row's model, effort, minimum version, and guard exactly. Synthesis and the per-finding gate stay in the main loop.
• Each critic gets the plan + its one lens + "find the strongest objection; default to finding a real problem", never the others' output. Inline any hard rule a critic needs — the subagent does not read your global config.
• Keep every lone-critic finding. A problem only one lens caught is the whole reason for the panel; never drop it as an outlier.
• Dedup across critics, rank by severity, present each as finding / evidence / suggested change, then use the exact per-finding gate from lib/gate-contract.md: Apply? [approve / edit / reject / skip]. Outside-voice findings are informational; the caller decides each.

Stop after the gated list. advisor returns judgment; it does not execute (that is sprint).