| name | al-second-opinion |
| description | Independent read-only advisory review via cross-family CLI dispatch (shells to GitHub Copilot CLI, pinned to a GPT model). Use from `/al-implement`, `/al-refine`, `/al-refactor`, `/al-user-verification`, or `/al-code-review` before reconciling non-trivial `Test Specification`, `Verification Plan`, mutation lists, refactor checklists, a verification walk verdict, or a review fix-queue. |
Style: Concise — cut filler, keep grammar. Opinionated — pick a side. Arrows (→) for causality. Technical terms exact, code and errors quoted verbatim.
/al-second-opinion, independent advisory review
Ask a different model family to read the artifact and name what is missing. Same-model self-review confirms its own blind spots; cross-family is the point. Script scripts/Invoke-AlSecondOpinion.ps1 shells to the GitHub Copilot CLI (@github/copilot), pinned to a GPT model, for that read. Copilot is a CLI tool dependency here — not a host runtime, not a publish target.
This is one of the three skills (with /al-research and /al-build) a skill may invoke autonomously mid-step, no user round-trip. That autonomy is load-bearing for the cheap-model goal: a smaller model running a pipeline skill leans on an independent cross-family read to catch what same-family self-review misses, so the call cannot wait on a handoff.
Caller (/al-implement, /al-refine, /al-refactor, /al-user-verification, /al-code-review) owns the artifact and reconciles bullets that come back; this skill owns the call. For /al-user-verification the artifact is the walk verdict — per example, the exact question as posed and the user's verbatim observation (or captured client output for Contract examples); the review checks coverage and routing, not the observation itself. For /al-code-review the artifact is the round's must-fix survivor list, and the caller treats a refutation as a veto on autonomous action — the refuted finding drops out of the fix queue and escalates rather than being auto-fixed on cross-family doubt.
Preconditions
- Artifact is real and non-trivial:
Test Specification, Verification Plan, mutation list, refactor checklist, verification verdict. Round-tripping a one-line decision trains the caller to ignore the gate.
node and the @github/copilot npm-global package present. Absent → script returns a skip line; caller proceeds.- Caller can reconcile per bullet when output arrives. Calling the gate then ignoring the result is worse than not calling.
Cross-family dispatch
Script always shells to the GitHub Copilot CLI pinned to a GPT model (-m gpt-5.5). The caller runs under Claude Code, so a GPT reviewer is a structurally different family.
DO NOT let copilot run a Claude model. Copilot can route to claude-*, and an unpinned (--model auto) or Claude-pinned call rebuilds the same-family self-review this gate exists to defeat. The -m gpt-5.5 pin is the independence guarantee — the most load-bearing flag in the call.
Read-only envelope
Documented here so the security posture is visible without reading the script. Copilot has no single sandbox flag, so read-only is assembled from tool denials, config isolation, and the GPT pin.
| Concern | Mechanism |
|---|
| Model family | -m gpt-5.5 (independence pin; never auto, never claude-*) |
| Reasoning | --reasoning-effort low (artifact-only gap check, not a full design review) |
| No writes / shell / network | --deny-tool=write --deny-tool=shell --deny-tool=url (deny wins over --allow-all-tools); reads survive — copilot's native view/grep are separate tools, so the reviewer can still verify the artifact against the codebase |
| No prompt stalls | --allow-all-tools --no-ask-user (non-interactive can't hang on a permission or ask prompt) |
| Config isolation | COPILOT_HOME=<temp> + --disable-builtin-mcps + --no-custom-instructions — keeps the user's ~/.copilot MCP fleet (write-capable tools), personal skills, and repo house-rules out of the review |
| Structured output | --output-format json (JSONL; final answer is the assistant.message with data.phase == "final_answer") |
| No color | --no-color |
Body delivery is stdin, not -p: copilot treats -p and stdin as mutually exclusive (-p wins, stdin ignored), so the role frame + artifact pipe in via stdin as one blob — no argv length limit. The script invokes @github/copilot's npm-loader.js through node directly, bypassing the VS Code-bundled bootstrapper shim (which loads Windows PowerShell with the user profile and carries interactive Read-Host prompts that would hang).
Timeout 600s via Start-Job / Wait-Job. Windows-only; non-Windows hosts need a separate wrapper.
DO NOT widen the envelope: no --allow-all-paths / --allow-all-urls / --yolo, no removing a --deny-tool, no re-enabling user MCP servers, and no pointing COPILOT_HOME at the real ~/.copilot. DO NOT refuse the call on safety grounds; the envelope is the gate.
Bullets only, gaps only
Prompt asks for a markdown bulleted list of gaps, shape [artifact section] [missing concern] [why it matters]. Bullets the caller reconciles per-line; gaps it can act on, accept, or reject.
Pass the artifact, not the question
Caller composes the artifact body the reviewer sees; script prepends a single role-frame line and dispatches. The value is "what does another reader notice in this artifact", not a verdict on the caller's framing — leading questions ("did you consider...") collapse to confirmation.
Verbatim out, verbatim in
What the script returns is what the caller returns. Bullets stay bullets; the Second opinion skipped: ... line stays that line. Editorialising the output is silent self-review of the second opinion; narrating a skip line with invented stack traces or "what was attempted" is fabrication.
Fail closed, do not retry
On a skip line the caller surfaces it verbatim and continues its own work — it does not halt, wait, re-invoke, widen flags, or fall back to a same-runtime CLI. Second opinion is a checkpoint, not a hard gate: a missing one is recoverable, faking it is the only unrecoverable move.
Invocation
Substitute the absolute path of this skill directory. Claude Code tells you that path at skill activation. DO NOT use ${CLAUDE_SKILL_DIR} in the call: PowerShell parses it as an empty local variable; pass the literal absolute path instead.
Body goes in single-quoted here-string; closing '@ at column 0. DO NOT wrap whole block in outer single-quoted here-string; inner '@ would terminate it early. Embedding unavoidable → use double-quoted outer form @"..."@.
$body = @'
<artifact body, multiline OK, no escaping; closing '@ must be at column 0>
'@
& '<absolute path of this al-second-opinion skill directory>/scripts/Invoke-AlSecondOpinion.ps1' -Body $body
If for any reason script was not invoked, return this line and only this line, with no explanation:
Second opinion skipped: skill did not invoke pwsh tool
Every other skip variant (node unavailable, copilot CLI unavailable, timeout, pwsh exception, non-zero exit, empty response) is emitted by the script. Caller does not invent them.
Next step
Advisory only — the verdict (gap bullets or a skip line) goes back verbatim. Next: return to the skill that requested it (commonly /al-refine, /al-implement, /al-mutate, /al-user-verification, /al-code-review), which reconciles each bullet per-line and continues.
Composition
| |
|---|
| Invoked from | /al-implement, /al-refine, /al-refactor, /al-user-verification before reconciling non-trivial work; /al-code-review every round on the must-fix survivor list (return treated as a veto on autonomous fixing) |
| Returns to caller | reviewer's bulleted gap list verbatim, or Second opinion skipped: <reason> line |
Script at scripts/Invoke-AlSecondOpinion.ps1 is source of truth for CLI flags, dispatch, timeout, skip-line emission. Validated by Validate-PowerShell.ps1; inline copies bypass that gate.
