Second Opinion

Second Opinion (Claude ⇄ Codex relay)

Consult the locally installed Codex CLI (codex exec, non-interactive) as an independent reviewer, then run a short debate. Both sides share their reasoning: Claude writes its thinking into the brief; Codex is asked to return a verdict plus reasoning. The goal is a better final answer, not agreement for its own sake.

When to use

• The user explicitly asks for a second opinion / Codex's view on something Claude produced or proposed (a generation prompt, an architecture, copy, a plan, a diff).
• High-stakes or taste-driven decisions where an independent model genuinely helps.

Do NOT invoke on your own for routine work; it costs a real Codex turn (~30–90s each round).

Setup facts

• Command: codex exec (non-interactive). Verified on codex-cli 0.144.x; flags are identical on Windows, macOS, and Linux.
• OS detection: read the platform from the session environment — win32 → use the PowerShell variants below; darwin/linux → use the bash variants. Never ask the user which OS they're on.
• Codex's default model + reasoning effort come from ~/.codex/config.toml (currently a GPT-5.x model at high effort). Only pass -m <model> if the user names a model.
• Codex loads the user's Claude skills from ~/.agents/skills at startup — you can tell it by name which skill to consult.
• A round's stdout transcript can be 100KB+ (Codex dumps files it reads). ALWAYS redirect the transcript to a log file and capture the final answer with -o.

Workflow

1. Assemble the brief

Create a working folder under the session scratchpad (e.g. <scratchpad>\second-opinion\). Write brief.md containing, in this order:

1. Role framing — "You are being consulted by Claude (another AI agent) for an independent second opinion. Do not rubber-stamp. Challenge weak choices. You may consult your available skills if relevant (e.g. <skill-name>)."
2. Task — what the user is trying to achieve, self-contained (Codex has none of the conversation context).
3. Claude's proposal — the full artifact under review (the prompt, plan, code, copy), verbatim.
4. Claude's reasoning — why these choices were made, including known doubts. This is the shared-thinking channel; be honest about weak spots.
5. Pointed questions — 2–4 specific things to challenge, not "thoughts?".
6. Response format — "Reply with a verdict line first: AGREE / AGREE WITH CHANGES / DISAGREE, then your revised version and your reasoning. Keep it under N words." (250 is a good default; caps keep rounds fast.)

Never put secrets, API keys, or credentials in the brief.

2. Round 1 — ask Codex

Windows (PowerShell):

$wd = "<scratchpad>\second-opinion"
Get-Content "$wd\brief.md" -Raw | codex exec -s read-only -C "<project dir>" --skip-git-repo-check -c model_reasoning_summary="detailed" -o "$wd\round1.md" - *> "$wd\round1.log"
Get-Content "$wd\round1.md" -Raw
Select-String -Path "$wd\round1.log" -Pattern "session id:"

macOS / Linux (bash):

wd="<scratchpad>/second-opinion"
codex exec -s read-only -C "<project dir>" --skip-git-repo-check -c model_reasoning_summary="detailed" -o "$wd/round1.md" - < "$wd/brief.md" > "$wd/round1.log" 2>&1
cat "$wd/round1.md"
grep "session id:" "$wd/round1.log"

• - reads the brief from stdin (avoids quoting issues with long briefs).
• -C <project dir>: point at the real project when the question references project files, so Codex can read them; otherwise point at the working folder.
• -s read-only always — Codex must never modify files; Claude implements any changes.
• Timeout: allow 300000ms; large briefs can take a few minutes.
• Save the session id from the log header — it's needed for relay rounds.
• If you want Codex's step-by-step thinking, skim round1.log (its intermediate messages appear after codex marker lines; skill/file dumps appear after exec lines — skip those).

3. Claude thinks

Read Codex's verdict and reasoning and genuinely evaluate it:

• Adopt what is actually better, and say so.
• Where you disagree, articulate WHY — a real technical or creative reason, not preference.
• Watch for factual corrections (Codex may know/verify things Claude assumed wrong) — verify against the relevant skill or docs when it matters.

4. Relay rounds (only if warranted, max 2 more)

Only continue if there is genuine disagreement or an open question. Write replyN.md with Claude's counterpoints and reasoning, then resume the SAME Codex session:

Windows (PowerShell):

Get-Content "$wd\reply1.md" -Raw | codex exec resume <session-id> --skip-git-repo-check -c model_reasoning_summary="detailed" -o "$wd\round2.md" - *> "$wd\round2.log"

macOS / Linux (bash):

codex exec resume <session-id> --skip-git-repo-check -c model_reasoning_summary="detailed" -o "$wd/round2.md" - < "$wd/reply1.md" > "$wd/round2.log" 2>&1

• resume inherits the original sandbox and workdir; no need to repeat -s/-C.
• In relays, ask for a per-point verdict: "CONCEDE / HOLD, one sentence of reasoning each, final version only if changed. Under 150 words."
• Stop when converged, or after round 3 total. Never capitulate just to converge, and never let Codex's confidence substitute for checking.
• Fallback if the session id was lost: codex exec resume --last ... (riskier — grabs the newest session globally).

5. Synthesize for the user

The final message must stand alone:

• The final artifact (revised prompt/plan/code) after the debate.
• What Codex agreed with, what it pushed back on, and what Claude changed as a result (with the why).
• Any unresolved disagreement: state both positions and Claude's recommendation.
• Where the round files live, in case the user wants the full transcript.

Modes

• Quick check (default when the user says "quick" or the stakes are low): one round, no relay, 150-word cap.
• Debate (default): up to 3 rounds as described.
• Deep: user asks for a thorough review — raise the word cap, let Codex read project files via -C, allow more rounds.

Failure handling

• codex not found → tell the user to install/check PATH.
• Auth errors → user runs codex login themselves (suggest ! codex login).
• Ignore startup noise in the log: MCP AuthRequired errors, skill-frontmatter load errors, "skills context budget" warnings — they don't affect the run.
• If a round times out, retry once with a longer timeout and a shorter brief before giving up; if Codex is down, say so and give Claude's own best answer.