Second Opinion

Problem with this skill? Run kendex report — it files to the owning repo automatically. Do not hand-file.

Cross-model second opinion via external AI CLI. The guarantee is that the model this session runs is never asked to second-guess its own work: every mode walks the SECOND_OPINION_MODELS roster in priority order and takes the first target that is available and runs a different model — Codex from a Claude Code session, Claude from a Codex session. When nothing eligible remains the run refuses and says why. Breadth is opt-in: SECOND_OPINION_COUNT raises the number of opinions a review collects, and with two or more the lanes run in parallel and the findings are unioned (contract below).

.agents/skills/second-opinion/scripts/second-opinion <mode> [options]

Workflows

Timestamp and agent are wrapper-stamped. In review and audit modes the wrapper overwrites the JSON timestamp field with its own UTC wall clock (date -u) after the model responds, so it records when the artifact was produced, never a model-serialized value. It overwrites agent the same way — external-<target> for a lane, external-union(…) for a union — so the marker is this skill's assertion of authorship rather than something a provider may omit or get wrong, and every artifact it writes is one its own cleanup can later recognize. Downstream freshness checks (orch review-artifact-check --file <path> [delegated_at]) validate filesystem mtime, and the stamped timestamp stays consistent with it.

Cross-model is enforced, in every mode. Identity is model-level and declared, not inferred from the CLI: this session's identity is SECOND_OPINION_CURRENT_MODEL when set, else the model the detected harness runs (claude, codex — the nearest harness ancestor in the process tree, then environment markers); each roster target's identity is SECOND_OPINION_<NAME>_MODEL, defaulting to its name. A target whose identity equals the session's is skipped loudly — --target and SECOND_OPINION_TARGET included — and a run with no eligible target exits 1 with every candidate and its reason, writing nothing and invoking nothing. A multi-model front end (Pi, OpenCode, Cursor) or an undetected harness with no SECOND_OPINION_CURRENT_MODEL has no identity to compare and is refused the same way — set it to the model the session runs, or to none when there is no session model (CI, a plain terminal), in that session's environment. A positively detected single-model harness beats any contradicting declaration, whatever its source. A claude or codex harness is direct evidence about this process; a declaration is a variable, and variables are inherited — a Claude session that exports one and starts a nested Codex session leaks it there. So a declaration that canonicalizes to a different identity than the detected harness is refused, naming both values and where the declaration came from; one that agrees changes nothing and is simply not used. Source decides only where detection cannot arbitrate (Pi, OpenCode, Cursor, undetected): there the declaration is the only identity available, and only a value in the session's own environment is a statement about this session — one from a project settings file (.env, kendex.settings.toml, .kendex/settings.toml, .env.local) is read by every session in the repo and is refused, naming the file. Declared values normalize from model ids, provider prefix and all, and surrounding whitespace is ignored (opus, claude-opus-5, anthropic/claude-opus-4, claude → claude; gpt-5.6-sol, openai-codex/gpt-5.6-sol → codex); a declared identity the roster does not spell is refused too — name the model in SECOND_OPINION_MODELS (a command is optional) or fix the declaration. A detected identity the roster does not name excludes nothing and the walk proceeds: the roster is the priority list of review targets, so naming only the cross-model one is a valid configuration.

Multi-lane review is opt-in breadth. With SECOND_OPINION_COUNT of 2 or more, review takes up to that many distinct eligible models from the roster, runs them in parallel on one pinned scope, and writes a single union artifact: findings deduplicated across lanes, per-lane provenance in qa_metadata.lanes, and — with --output — each lane's own artifact kept beside the union as <output>.<target>.json. Lane resolution, merge rules, artifact placement and permissions, scratch durability, and the failure taxonomy: references/multi-lane.md.

The review prompt reads the repo's own instruction files. review mode appends, when present, the files matched by the SECOND_OPINION_REVIEW_INSTRUCTIONS glob list — default AGENTS.md, review-bots.md, .github/instructions/*.instructions.md, .github/copilot-instructions.md, resolved inside --cwd — so repo-rule adherence is reviewable locally, same as GitHub bots. When the list carries the literal AGENTS.md entry, nested AGENTS.md files governing the changed paths are collected too (ancestor-directory walk, same containment and dedupe). Set the variable empty to disable. Matches must be regular files physically inside the reviewed repo: a committed symlink (the file or a parent directory) pointing outside --cwd is skipped loudly, never followed — the reviewed checkout is untrusted and must not be able to leak host files into the prompt. The prompt reviews through explicit lenses (correctness, security/fail-open, adversarial inputs, Bash-3.2/portability, repo-rule adherence, docs-vs-code drift, test adequacy); only pure style and naming opinions stay out of scope.

The artifact records the reviewed head. review mode stamps qa_metadata.reviewed_head — the reviewed range's endpoint commit, captured when the scope is derived (before the model runs), never re-sampled from a mutable HEAD afterward — so callers can budget review passes per pushed head — a new head is a new round — instead of per submission.

Review scope is wrapper-derived. In review mode the wrapper derives the scope from the worktree before invoking the external CLI — current branch, diff range (--range or origin/BASE...HEAD), diffstat, and the changed-file list are embedded in the prompt, so the external model is never asked to guess its own scope. A first response that is unparseable, or structurally incomplete, gets one retry resending the full original request (scope block included) alongside the captured response, so the retry session reviews the same scope instead of answering context-free. Unusable responses never become the artifact — see Error Handling. orch review-artifact-check independently rejects self-reported no-review artifacts (reason no_review) and qa-shaped artifacts missing their finding arrays (reason incomplete), regardless of verdict.

Common Options

All modes accept:

Execution Rules

• Execute all workflow sections in order. The workflow decides what to skip via "Skip if" conditions — never skip based on your own scope assessment.
• <output_format> tags are literal templates: fill [PLACEHOLDERS], omit empty lines, add nothing else, do not paraphrase.
• Pass --target when the user explicitly requests a specific model/CLI (e.g., "use Claude", "ask Codex"). Otherwise omit it — the script selects from the roster and the current session's model. A forced target that runs this session's model is refused; report the refusal, do not work around it.
• Do not pass --timeout unless the user explicitly asks for a different value for this specific call — the script reads the default from project config.
• Always pass --cwd with the absolute project root path. Never use --cwd . — the external CLI needs the full path to find project files.
• For quick mode, you can pass the question as an inline argument instead of writing a file: second-opinion quick "your question here" --cwd /path

Configuration

Set non-sensitive defaults in kendex.settings.toml under [env]. Existing .env.local and .env values still work; .env.local wins. The one exception is SECOND_OPINION_CURRENT_MODEL — export it in the environment of the session that needs it; a value in any project file (.env, kendex.settings.toml, .kendex/settings.toml, .env.local) is refused.

Project installs seed kendex.settings.toml from this skill's kendex.settings.toml.example when missing and merge only absent second-opinion keys into existing files.

Default commands

Claude (identity claude):

SECOND_OPINION_CLAUDE_CMD="claude -p --no-session-persistence --model opus --effort max --allowedTools Bash(read-only:true),Read,Glob,Grep"

Codex (identity codex):

SECOND_OPINION_CODEX_CMD="codex exec -m gpt-5.6-sol -s read-only -c model_reasoning_effort=xhigh --ephemeral"

To customize, copy the full command into kendex.settings.toml for shared defaults or .env.local for personal overrides and edit any flags. The entire variable is used as-is.

Error Handling

On script failure (non-zero exit), stderr contains a JSON error object:

{"error": "description", "target": "codex"}

No exit leaves a previous run's output behind. Every mode that writes --output — review, audit, challenge, quick — clears that path as soon as the arguments are read, together with exactly the sidecar records that mode can produce: review and audit add .raw.txt, .retry.txt, .failed.json, .noreview.json and .incomplete.json, while challenge and quick preserve no such record and so clear their output path alone — beside their --output those five names are your files: ahead of a bad-flag error, the jq check, the mode check and --timeout validation, and before target selection or any CLI run. So a refusal, timeout, or unusable response cannot be read as this run's result by a caller that continues past the advisory non-zero exit. --output is a path the caller designated for this run's output, so no question of authorship arises there; --help and detect never reach the write and clear nothing.

One rule governs every deletion: a path is removed unconditionally only when this run will write it; anything else must first prove the skill wrote it — the review schema plus the external-… agent marker. Two paths meet the first half: --output itself, and each <output>.<lane>.json a multi-lane run is about to write. Every other sibling goes through the ownership check, including roster-named ones — at the default SECOND_OPINION_COUNT=1 a run writes no lane file at all, so <output>.codex.json holding your data survives while the same name carrying the skill's marker is reclaimed. That check parses the candidate, so it requires jq: without it the designated output is still cleared (a path the run writes needs no proof), but sibling lane artifacts cannot be shown to be ours and are left alone — the run says so on stderr before exiting on the missing dependency. A sibling sharing only the name shape (<output>.notes.json), or sharing the schema under another agent (an internal reviewer-* artifact), is never touched.

Options take their value in either form, --opt VALUE or --opt=VALUE; the split form refuses a value that begins with - (it would otherwise swallow the following flag), so a dash-leading path is passed as --output=-report.json.

Runtime failures emit the JSON error object above; pre-flight configuration errors — unknown flag, an option missing its value, missing mode, non-integer --timeout or SECOND_OPINION_COUNT, missing jq, a directory or an uncreatable parent at --output, a missing --cwd — are plain Error: lines on stderr with exit 1. They are plain by necessity: they can precede the jq dependency check, which is what would format the JSON.

Multi-lane review maps lane failures into the same contract: a failing lane is recorded inside the union artifact (qa_metadata.lanes, coverage: "degraded") with the run still exiting 0; only when every lane fails does the run exit 4 (at least one lane answered unusably) or 5 (no lane answered), writing no artifact.

If the script fails during the orch review-pr or submit-pr (local pre-PR review) workflows, continue — external review is advisory.