Advisor

Advisor (GPT-5.6 consulted)

Consult an independent GPT-5.6 reviewer for one specific decision point, at the same reasoning effort you are currently running at. This is the Codex-side counterpart of the Claude-side advisor skill (plugin/skills/advisor/ in open-science-skills), built the same session for the same reason: a single-turn advisory consult, not a delegated implementation.

Which 5.6 tier, and why it matters. There are three distinct GPT-5.6 tiers, not one model: gpt-5.6-sol (flagship, "most capable"), gpt-5.6-terra (balanced, "everyday work"), gpt-5.6-luna (fast/cheap). This skill defaults to gpt-5.6-terra for routine consults — it is cheaper and confirmed reliable. gpt-5.6-sol is confirmed working on both hosts this repo runs on (July 2026, both authenticated via ChatGPT login) — an earlier "rejected outright under ChatGPT-account auth" finding no longer reproduces; whether that was a temporary rollout gate or a stale/outdated-CLI artifact is unclear, so re-verify with a smoke test if sol ever errors on your account rather than assuming it's gated. Pass --model gpt-5.6-sol explicitly for a stronger, costlier consult — see Effort calibration below.

Read scripts/sol-advisor.sh before the first run — the flags and the effort-handling note there are load-bearing.

Sandbox constraint — read this before your first run

sol-advisor.sh shells out to a nested codex exec process. Confirmed by direct reproduction (July 2026, both hosts this repo runs on): a codex exec process running under any sandbox mode (read-only or workspace-write) cannot spawn a working nested codex exec child — it fails immediately with Error: failed to initialize in-process app-server client: Operation not permitted (macOS) or Read-only file system (Linux). This is structural — the OS sandbox applies transitively to the whole process tree — and is not fixed by passing --dangerously-bypass-approvals-and-sandbox to the nested call itself; only an unsandboxed top-level session avoids it entirely.

• If you are interactive (not codex exec invoked non-interactively) and this failure surfaces, request escalation for that one shell command (sandbox_permissions: require_escalated, with a short justification) so it runs outside the sandbox — this is exactly the case the built-in escalation path exists for.
• If you are running non-interactively (codex exec at the top level shows approval: never in its own banner — no human is available to approve an escalation), this call cannot succeed. Report the failure honestly; do not silently produce the review yourself in place of the consult (that defeats the point of an independent second opinion) and do not retry indefinitely.

When to use

• Before substantive work — before writing, before committing to an interpretation, before building on an assumption. Orientation (finding files, reading a source) is not substantive; writing, editing, and declaring an answer are.
• When you believe a task is complete. Make the deliverable durable first (write the file, save the result, commit the change) — the consult takes real time.
• When stuck — errors recurring, an approach not converging, results that do not fit.
• When considering a change of approach.

On tasks longer than a few steps, consult at least once before committing to an approach and once before declaring done. Give the advice serious weight: if a step you followed on the consult's advice fails empirically, or you have primary-source evidence contradicting a specific claim, adapt. If your own evidence and the consult's advice point different directions, surface the conflict in one more consult rather than silently picking a side.

Effort calibration — read this before running

The consulted model should run at the same reasoning effort you are currently running at, not a hardcoded default. Codex does not expose its own live reasoning-effort setting as an environment variable inheritable by subprocesses (confirmed empirically: dumping env inside a running codex exec call shows no CODEX_REASONING_EFFORT or equivalent — only Claude Code's own $CLAUDE_EFFORT leaked through from an enclosing Claude Code process, which is not your effort level when you are the one running as Codex).

What Codex does expose: the "reasoning effort: <level>" line shown in your own session banner at startup, visible in your own context. Before running the consult, read that value from your own session context and pass it explicitly:

scripts/sol-advisor.sh --prompt-file <briefing-path> --out <output-path> -C "$PWD" --effort <your-own-level>

Do not omit --effort and let it silently default to the script's fallback (high) unless high genuinely is your own current level — the fallback exists so the script never errors on a missing flag, not as a substitute for checking your own level.

Steps

1. Compose the briefing. A self-contained document: the task, what you have done so far (key steps, findings, decisions), your current approach or the specific claim to check, and the precise question. Concrete — file paths, line numbers, the actual claim. Save to a scratch file. The consulted model has no access to your conversation beyond this file.

2. If declaring a task complete, make the deliverable durable first.

3. Read your own current reasoning-effort level from your session context, then run:

   scripts/sol-advisor.sh --prompt-file <briefing-path> --out <output-path> -C "$PWD" --effort <your-level>
   

4. Read the output file and weigh the advice. Integrate it; if you diverge from it, be able to say why.

Notes

• scripts/sol-advisor.sh --check verifies the codex CLI is on PATH.
• The spawned session runs --sandbox read-only and --ephemeral — advisory only, no edits, not saved as a resumable session.
• Default model is gpt-5.6-terra, not gpt-5.6-sol — cost, not reliability. There are three distinct GPT-5.6 tiers (sol flagship / terra balanced / luna fast) — not one model with a shorthand. An earlier finding claimed gpt-5.6-sol was rejected outright on a Codex CLI authenticated via codex login with a ChatGPT account (as opposed to an API key) — "The 'gpt-5.6' model is not supported when using Codex with a ChatGPT account." This has since been re-tested (July 2026) and no longer reproduces: sol now works on hosts authenticated the identical way (codex login status → Logged in using ChatGPT). Whether the original block was a temporary rollout gate or an artifact of an outdated CLI on that host is unclear — if sol ever errors on your account, check codex --version first (an outdated CLI rejects even terra/luna with a different, version-specific error) before concluding it's an account gate. Pass --model gpt-5.6-sol explicitly for a stronger reviewer at higher cost. If the model you asked for is unreachable for any reason, report it and ask whether to stop or use a named replacement — do not silently fall back further (e.g. to gpt-5.5) for what is supposed to be a genuinely independent second opinion (a reviewer identical to the executor defeats the point of this specific skill; if a same-model consult is actually fine for the task at hand, the calling agent should use its own judgment directly rather than spend a consult on it).
• Effort enum: none, minimal, low, medium, high, xhigh (Codex's scale — no max; Claude Code's max has no Codex equivalent).
• Companion skill: plugin/skills/advisor/ (Claude-side) — same pattern, consulting Fable 5, effort read from $CLAUDE_EFFORT instead of the session banner.