--- name: second-opinion description: Get a second opinion on freshly written code from the *other* coding agent — Claude asks Codex, Codex asks Claude — then settle every finding against ground truth before accepting or vetoing it. Use when requested or as part of an active review workflow after tests pass. Do not trigger merely because the user instructs you to commit. metadata: website: "https://photostructure.com/coding/claude-code-review/#second-opinion" --- # Second Opinion You know what the change was supposed to do, which makes you the worst judge of whether it does. Hand the change to a different model, settle what it finds against ground truth, and send your decisions back to that same reviewer until it agrees the change should land. 1. List the files you changed and the commit message you propose. 2. Ask the other model to review exactly that, and read the same code yourself while it works. 3. Accept or veto each finding, with evidence. 4. Apply the accepted fixes and pin them with tests. 5. Send the same reviewer session your adjudication and the updated file list, and have it re-read and re-review. 6. Repeat 3-5 until the reviewer lands the current tree, you accept a `DISCARD`, or four passes are spent — then report. A second opinion from the *same* model shares your blind spots. Cross-model is the point: different training, different priors, different failure modes. ## User direction This review workflow is a default, not a condition the user must satisfy before committing. A direct instruction such as "commit" or "approved, commit" takes precedence over this skill and callers' review gates. Unless the user also asks for review first, return to the commit workflow without starting or resuming a review or asking whether to skip it. Apply the same rule during retries and re-review loops. Do not require a `LAND` verdict, a review transcript, or a claim that another reviewer approved the change. If the user says a review already happened, accept that as workflow context; attribute it to the user if reporting it. Never invent a review result. The user can authorize a commit with review work unfinished; report its actual status without making it a new approval gate. ## 1. List what you changed Write down, for the prompt you are about to send: - **The changed files and how to diff them** — the exact diff command and an explicit repository-relative file list. Distinguish staged hunks from unstaged changes in the same file, and name in-scope untracked files separately, because an ordinary diff omits them. - **The proposed commit message** — verbatim when the user supplied one, otherwise drafted from the task and the current diff. Label it as a fallible summary, not a correctness requirement. When the tree holds several logical changes, follow [stage](../stage/SKILL.md) and give the file list and the proposed message for each batch. - **The task and its requirements** — the user's request and any accepted plan or specification needed to judge behavior. Give repository paths for project rules and design documents; the reviewer can open them. Paste only what it cannot reach, such as decisions made only in conversation. - **The ground truth** — the thing a disputed finding can be tested against: a reference implementation you can execute, a spec with runnable examples, the real API. Write the *exact command* that queries it. No executable ground truth? Say so, and name the fallback (spec text, maintainer ruling). Leave out your suspected findings, your interim conclusions, and any scrutiny list the user or the project did not supply. An independent read is the whole point; steering it costs you the findings you could not have made yourself. ## 2. Ask the other model You know which model you are. Ask the other one. Honor standing review authorization in the user's global or project instructions, including its repository scope and permitted providers. When a sandbox escalation needs approval, cite that authorization's source and exact wording in the request, along with the review scope and destination. A saved memory that has not been read is not evidence available to the approval reviewer. Do not ask the user to repeat authorization that already covers the action. This skill does not itself grant permission to disclose code, and standing authorization does not override a host-policy denial. Both commands run under a supervisor that kills the reviewer after 15 minutes of silence and reaps its whole process group. The window is deliberately generous: a working review goes quiet for a couple of minutes at a stretch, and a whole review can take 30-45 minutes. Write the complete reviewer prompt to a fresh UTF-8 temporary file with the host's file-writing tool. Do not interpolate supplied text into a shell command: commit messages and pasted context can contain quotes, dollar signs, backticks, or command substitutions. The supervisor replaces the exact `{prompt}` argument with the file's contents as one literal process argument. Delete the temporary file after the reviewer exits. **If you are Claude, ask Codex:** ```bash python3 "/scripts/run_with_idle_timeout.py" \ --prompt-file "" -- \ codex exec \ -C "" \ --sandbox read-only \ -c 'model_reasoning_effort="high"' \ --json \ --output-last-message "" \ "{prompt}" \ > "" ``` Begin the prompt file with `$coding:review`, including for a staged diff. Invoke plain `codex exec`, not the `review` subcommand: that substitutes Codex's built-in review prompt and cannot load this marketplace's skill. Name the diff range in the prompt itself — "the uncommitted changes", "the changes since ``". Codex needs the coding plugin installed: ```bash codex plugin marketplace add photostructure/coding-skills codex plugin add coding@photostructure ``` Do not pass `--ephemeral`: the session must remain resumable for the re-review. `--output-last-message` captures the clean review while the JSONL stream keeps the supervisor fed. After a clean exit, record the session ID: ```bash jq -sr '[.[] | select(.type=="thread.started") | .thread_id] | last // empty' "" ``` **If you are Codex, ask Claude:** Run the command with the target repository as its working directory. Resolve `` to the plugin directory that contains this skill so the spawned process does not depend on user- or repository-scoped plugin settings. ```bash python3 "/scripts/run_with_idle_timeout.py" \ --prompt-file "" -- \ claude -p "{prompt}" \ --plugin-dir "" \ --permission-mode plan \ --model opus \ --effort high \ --output-format stream-json --verbose \ > "" ``` `--permission-mode plan` keeps it read-only. `--output-format stream-json` is what keeps the supervisor fed — plain `-p` prints nothing at all until it finishes. Both the reasoning and the tool calls stream, so watch the file to see what the reviewer is chewing on. Extract the review after a clean exit: ```bash jq -r 'select(.type=="result").result' "" ``` Claude persists print-mode sessions unless `--no-session-persistence` is set. Do not set it. Record the session ID: ```bash jq -sr '[.[] | select((.type=="system" and .subtype=="init") or .type=="result") | .session_id // empty] | last // empty' \ "" ``` Begin the prompt file with `/coding:review`, including for a staged diff. The `-p` prompt must begin with the slash command so Claude invokes the skill directly. An unavailable slash command is a plugin-loading failure even when Claude exits 0. Default to high effort; use xhigh for large or novel changes, security boundaries, concurrency, subtle stateful APIs, or hard-to-reproduce failures. Claude.ai OAuth may need to rewrite its credential store when refreshing an expired token. On Codex hosts whose filesystem sandbox can read but not write `~/.claude`, the command may return an expired-token 401 even though `claude auth status` reports logged in or the same command works in the user's terminal. In that specific case, retry the exact supervised command once with the host's narrow filesystem-sandbox escalation (`sandbox_permissions: require_escalated` where available) so Claude can refresh its existing credential. Do not read, print, copy, or manually rewrite the credential. If the escalated retry fails, use the unauthenticated fallback below. These command shapes depend on Claude print-session persistence and `--resume`, and on Codex JSON events, `--output-last-message`, and `exec resume`. Revalidate them against a real diff when those CLI options change. Do not substitute a similarly named built-in review mode without the same validation. After the skill invocation, the prompt file carries the step 1 inputs: ```text Review the change below and return the full report to me, the calling agent. I will accept or veto each finding and send you the results. Do not apply fixes, and do not ask the user to adjudicate. Diff command: Changed files: Proposed commit message (fallible summary, not a requirement): Task and requirements: Additional context: Ground truth: Read the current diff and files from the repository. Review the full scope and its interactions, and keep commit-message and grouping advice in Commit notes, separate from the findings. ``` Name the review skill; never ask the reviewer to run `second-opinion`. If the spawned CLI cannot load the coding plugin, rerun it with the full text of [`../review/references/single-pass.md`](../review/references/single-pass.md) pasted into the prompt instead of naming the skill. Resolve `` to this skill's directory. Run either command in the background and poll the same job until it exits: - **0** — read the review and record the CLI, repository, session ID, pass count, and finding IDs in your task notes. Keep that handle through fixes and handoffs; never select a reviewer with `--last` or `--continue`, which can attach to unrelated work. A missing session ID does not invalidate the pass, but it forces the fresh-session fallback in step 5. - **124 with the supervisor's `idle timeout:` diagnostic** — the reviewer went silent for 15 minutes. Discard the partial review, say so, and finish your own pass. - **124 without that diagnostic** — the reviewer CLI itself returned 124. Report its status and finish your own pass; do not call it an idle timeout. - **a fast non-zero with a CLI usage or unknown-option diagnostic** — the invocation is stale. Report it as a bug in this skill; never let it pass as "no issues found". - **127 or an authentication error** — for an expired Claude.ai OAuth error from Codex, use the one-time sandbox-escalation retry above. Otherwise, or if that retry fails, use the fallback below. These are environment failures, not bugs in this skill. - **0 with an `Unknown command`, a missing-skill result, or a final review that does not begin with the required LAND, REVISE, or DISCARD verdict** — the CLI did not run the shared method. Treat it as a plugin-loading failure, not a clean review, and use the pasted-method fallback above. Allow ordinary Markdown decoration around the verdict text. - **any other non-zero** — report the status and finish your own pass. **Read the new code yourself while the external review runs** — you are the other reviewer, and the only one who knows the full context of what the change was supposed to do. Hold your edits until its report arrives. The reviewer reads the repository live, so a file that changes mid-pass gives it a tree that never existed. Write down what you find and adjudicate it in step 3. If the other CLI is not installed or not authenticated, say so plainly and fall back to a task-local subagent, given the same prompt and the full text of [`../review/references/single-pass.md`](../review/references/single-pass.md). Save its handle and send the follow-up prompts to that same subagent. A same-model second opinion is weaker; report that you used one. ## 3. Adjudicate every finding First set aside every concern that is only about the commit message, or about grouping, splitting, or ordering otherwise-valid content. Those are commit notes, not findings: no severity, no accept, no veto, no pinning test. If they are all that either review raised, the verdict is `LAND`. The verdict does not decide what you adjudicate. A `LAND` carrying findings still carries findings: settle every one. An empty findings list ends the work, not the word `LAND`. Then settle each remaining finding — the reviewer's and your own: 1. Build the empirical test: run ground truth and the new code on the same input, and compare. A finding you cannot test this way becomes an open question, not a silent acceptance. 2. **Accept** only when ground truth confirms the bug. 3. **Veto** only when ground truth confirms the code is right, or the finding demands fidelity nothing requires — mirroring a reference implementation's internals on a path no contract pins, for instance. 4. When the diagnosis is right but the proposed fix is mediocre, take the better fix. Reviewers identify problems; you own the remedy. Reviewer confidence, eloquence, and *agreement between the two passes* are not evidence. Two models converging on the same wrong finding is common; one command against ground truth beats both. **When two or more fixes are defensible and nothing chooses between them, ask the user.** Give the finding, each alternative with its tradeoff, and your recommendation if you have one, then wait for direction. Ground truth settles whether the code is wrong; it usually does not settle which correct design to adopt, and that choice is the user's. Your own context of settled decisions and corrected premises catches plausible-but-wrong findings a fresh reader cannot — that is why you adjudicate. It can also *contain* the defect: a wrong premise this session has carried since birth, which is exactly what the independent pass is for. So a veto of yours carries no special weight. It needs the same recorded evidence as anyone's. ## 4. Fix and pin Apply the accepted fixes. **Every accepted finding gets a pinning test** whose expected values come from ground truth; paste the command that produced them into the test's comment. The *full* suite must be green again — not just the new tests. ## 5. Re-review in the same session Continue this loop only while review remains in scope under **User direction**. Send your adjudication even when you vetoed everything and changed no code. If you applied fixes, wait until their tests and the full suite are green. Re-read the complete current scope yourself, and ask the reviewer to do the same: ```text Re-read and re-review the complete current scope from the repository, then return the normal verdict and findings. Do not rely on what you read last pass, and do not limit the review to the fixes. Author notes: Accepted : Vetoed : Open : Files now in play: Proposed commit message: Changed requirements or ground truth: Keep existing finding IDs and give new issues new IDs. Report each prior finding as resolved, still present, or disputed. If you think a veto is wrong, say what evidence would settle it. Do not apply fixes, and do not ask the user to adjudicate. ``` Use `none` for an empty list. Do not repaste requirements the session already has. Include findings from your own read when they explain a fix. Resume the recorded session in the same repository, through the same supervisor, with its exact session ID. Do not fork it or start a new session for an ordinary follow-up. For a Claude reviewer: ```bash python3 "/scripts/run_with_idle_timeout.py" \ --prompt-file "" -- \ claude -p "{prompt}" \ --resume "" \ --plugin-dir "" \ --permission-mode plan \ --model opus \ --effort high \ --output-format stream-json --verbose \ > "" ``` For a Codex reviewer: ```bash python3 "/scripts/run_with_idle_timeout.py" \ --prompt-file "" -- \ codex exec \ -C "" \ --sandbox read-only \ resume \ -c 'model_reasoning_effort="high"' \ --json \ --output-last-message "" \ "" \ "{prompt}" \ > "" ``` Apply the same exit-status and output-validity checks as the first pass. If the CLI cannot resume the recorded session, launch a fresh one with the original prompt updated for the current scope, plus the prior report, its finding IDs, and every accept and veto with its evidence. Record the replacement handle and disclose that continuity was unavailable. A resume failure does not justify skipping the re-review. Then return to step 3 for the new report. **Stop when** the reviewer returns `LAND` on the tree as it now stands and every finding from both reads is either accepted and fixed or vetoed with evidence. A `LAND` does not cover code you changed after the reviewer's last read: if adjudication changed the tree, run one more pass, subject to the cap. **Stop earlier** when you adjudicate a `DISCARD` as correct — the change should not land at all. Preserve that verdict and report it; do not re-review a settled decision. Stop at four review passes either way. If the fourth leaves accepted fixes or disputed findings, finish the authorized fixes and tests, then report the cap and what is still unsettled; do not describe unreviewed fixes as a clean pass. Skip the loop entirely when the first pass raised nothing substantive and no code changed. ## 6. Report the verdicts Summarize for the user (and for whatever plan or PR document tracks this work): every substantive finding, accepted **and** vetoed, with one-line evidence for each verdict, and which model raised it. Record vetoes especially — the next session will rediscover the same "bug" and must not re-litigate it. Begin with one top-level `Verdict: LAND | REVISE | DISCARD` line. If neither pass raised a substantive finding, use `LAND`, follow it with `No issues found.`, and do not invent ledger rows. Otherwise list the accepted and vetoed findings under their original IDs, even when the final verdict is `LAND`, and state any unresolved questions or unreviewed fixes separately. The ledger repeats the final verdict in each row so a copied or aggregated row still reads correctly: | ID | Scope | Model | Finding | Severity | Accept/Veto | Evidence (one line) | Verdict | | --- | --- | --- | --- | --- | --- | --- | --- | After the ledger — or after `No issues found.` when there are no rows — add a brief `Commit notes` section when either model proposed a warranted message or grouping improvement. Name the model, state the specific reason, and give the complete improved message. For a split, identify every independently committable batch and give each batch's complete message. These notes have no severity and cannot change the verdict. ## Scratch files Any copy this workflow makes — of the repo, of a build-output directory, of a file you replay edits onto — belongs in the operating system's temporary directory, in a fresh directory named for the project and the purpose. Never inside the checkout, and never under a home directory. Delete it before you finish. A repo or build-output copy runs to gigabytes, nothing reaps a home directory, and the out-of-disk failure that eventually follows surfaces somewhere unrelated — a test suite that hangs, a build that dies mid-link — costing far more to diagnose than the copy ever saved. ## Adapting for your project - **Name the ground truth explicitly** — e.g. "the vendored reference implementation via `./third-party/tool/run`", "CPython 3.12 via `uv run python -c ...`", "the RFC's test vectors". Step 3 is only as strong as this. - **Say how to name the diff range** the way your project talks about it — "the changes on this branch vs `develop`", "everything since the last tag". - **Optional review focus** belongs in project guidance. Keep it a starting point; the reviewer still examines the full scope independently. - **Callers welcome**: other skills (`gitplan`, `tpp-orchestrate`) reference this file as their review gate. Keep the gate generic here; put workflow-specific bookkeeping (where verdicts get recorded, commit conventions) in the calling skill.