advisor-fallback

name	advisor-fallback
description	Client-aware manual tiered-up adversarial advisor (claude subagent / codex exec), fired at the natural discretionary advisor points

Advisor Fallback (manual tiered-up advisor)

Purpose

The reliable advisor is a manual tiered-up adversarial reviewer you invoke yourself — not the advisor tool / /advisor slash command, which is flakey and, for codex, does not exist. This skill is the standard advisor path for both clients: at the moments you would normally consult an advisor, dispatch a fresh, blind, one-tier-up reviewer over the work and weigh its findings.

This is the same role the advisor tool plays, delivered by a subagent (claude) or a codex exec review (codex) so it works regardless of which client you are.

When to fire it (discretionary — NOT every time)

Fire the advisor at the same natural, judgment-based points a normal advisor call would happen — it does NOT fire on every task or every review gate:

• Before committing to an approach or interpretation (before substantive work).
• When stuck — recurring errors, an approach not converging, results that don't fit.
• When you believe the work is complete, before declaring it done.

Skip it for trivial mechanical steps. This is your judgment, exactly as a normal advisor is consulted by judgment, not on a fixed cadence.

How to invoke it — pick your client's branch

Resolve the one-tier-up reviewer model from your OWN client's ladder, then run that branch. Both branches are report-only: the reviewer identifies problems with evidence; it does not rewrite the work.

If you are a Claude session

Dispatch a fresh, blind subagent with the Agent tool (subagent_type="general-purpose", model=<one-tier-up>), giving it the task, the change/decision, the evidence, and the specific questions to pressure-test. Have it follow the ironclaude:adversarial-review methodology (orientation → deep read → verify every finding with grep/read evidence, drop the unverified → severity-classified, report-only).

Claude tier ladder (one tier up): sonnet → opus, opus → fable, fable → (top tier — skip; no higher advisor). If the resolved model is fable and Fable is unavailable, use opus (read the flag at IRONCLAUDE_FABLE_STATE_PATH if set else ~/.ironclaude/state/fable_unavailable.json; on any read error treat Fable as available). This is the existing behavior — model=fable, else model=opus.

If you are a Codex session

Run a one-tier-up codex exec adversarial review over the work. Pass the code/diff to review INLINE in the prompt — do not rely on the reviewer grepping the repo (see the gotchas). Command (live-validated 2026-07-22 — the grader codex exec family minus --output-schema):

codex exec --json --ephemeral --skip-git-repo-check -s read-only -m <ONE_TIER_UP_MODEL> -

• Deliver the adversarial-review instruction + the code/diff under review on stdin (trailing -).
• Read the findings from the last item.completed event whose item.type == "agent_message", field .text (the model may emit intermediate narration; take the last agent_message). Without --json the response text is printed directly — either form works; --json + last-agent_message is the reliable parse.
• Auth is your ChatGPT subscription (codex login status → "Logged in using ChatGPT") — never an API key.
• Ask the reviewer to follow the ironclaude:adversarial-review standard (verify findings with evidence, report-only, severity-classified).

Codex tier ladder (one tier up): luna (haiku) → terra (sonnet) → sol (opus). Codex has no fable tier, so at the ceiling sol → sol: run a same-tier BLIND pass (a fresh blind review at the same model still catches defects — most advisor value is blindness, not tier). The reviewer model stays a codex model — codex is its own advisor peer; do not cross to a claude model.

Codex gotchas (from the 2026-07-22 live probe)

• The ephemeral codex exec reviewer fires the IronClaude hooks and lands in professional-mode undecided, so its Bash is guard-blocked. This is why you pass the review content INLINE: the reviewer reasons over what you give it (and can still do non-bash file reads). Report-only review of inline content is unaffected.
• Session litter: --ephemeral is codex-session-ephemeral, not IronClaude-session-ephemeral. Each invocation leaves an undecided/idle row in ~/.claude/ironclaude.db (sessions) and an ~/.claude/ironclaude-session-<pid>.id file. These are inert to active workflows; be aware they accumulate (a periodic undecided/idle sweep is a reasonable follow-up). The same applies to the codex grader.

Weigh, don't obey

Treat the reviewer's findings as an adversary's evidence, not authority. Verify each material finding against the current source; reconcile conflicts with evidence. "No advisor available" never means "proceed unreviewed" — it means run this manual reviewer instead.