opus-advisor

name

opus-advisor

description

Consult an independent read-only Claude Opus advisor when a decision can affect security, privacy, money, data loss, deployment, or public API compatibility; when investigation leaves two plausible directions with materially different consequences; or when a consequential completion claim depends on an unverified assumption or critical behavior that normal tests cannot cover. Also use when the user requests an independent or Opus review. Skip routine edits and uncertainty that local inspection, documentation, or tests can resolve cheaply. Opus challenges Codex's reasoning and evidence; Codex retains responsibility for edits and the final judgment.

Opus Advisor

Use a fresh Opus context as an independent advisor. Opus inspects the project and returns its evidence, domain knowledge, objections, and verification demands; Codex performs the edits and owns the conclusion.

Choose one mode

• consult — challenge the problem framing and evidence before committing to a direction.
• review — challenge work in progress or its correctness argument.
• gate — challenge the finished work and verification evidence before claiming completion.

Combine overlapping needs into one call. For example, one gate call can cover correctness verification and the final audit.

Prepare the request

Run the launcher from the project under review:

<skill-directory>/scripts/ask-opus <consult|review|gate> '<request>'

Run ask-opus outside the active Codex sandbox with escalated execution since it require network outbound to Anthropic. Outer escalation lets the Claude CLI use the host network and configured proxy; it does not replace or relax the launcher's own read-only sandbox for Opus tools. If policy denies the required retry, report the advisor as unavailable.

Resolve <skill-directory> from this loaded SKILL.md; do not assume the skill is inside the current project.

The launcher grants read tools host-wide access and confines Bash writes to an ephemeral scratch directory with network access blocked. It removes that directory after the call and times out after 900 seconds by default; set OPUS_ADVISOR_TIMEOUT_SECONDS when a different bound is justified. Progress is streamed to stderr; stdout contains only the final advisory report.

Host-wide reads can expose source and system files to Claude. Do not invoke the advisor where that visibility to the model provider is unacceptable. Any tool permission denial invalidates the advisory call; the launcher fails closed instead of returning a partial report.

External disclosure must be authorized by the active execution policy or the user; this skill does not grant that approval. When official Anthropic services are an approved recipient, this resolved launcher is a subset of that provider-level trust. Keep secrets out of the request text and do not direct Opus to credential stores: the launcher clears alternate-provider and custom endpoint routing, denies credential environment variables to sandboxed tools, but does not restrict file reads. It retains standard proxy and TLS environment settings needed for host connectivity; those settings remain part of the disclosure boundary.

Give Opus:

• the goal and success criteria;
• relevant constraints and artifact paths;
• the candidate decision, plan, or implementation when one exists;
• commands and observed results already used as evidence;
• any context that exists only in the Codex conversation;
• the precise uncertainty or verdict requested.

Ask open questions when seeking tacit domain knowledge. Avoid seeding the expected answer. Let Opus inspect the workspace instead of pasting a large synthetic summary when the evidence is locally available.

Integrate the advice

Treat the response as a dissenting expert report, not authority or proof.

• Verify workspace claims against files, logs, or measurements.
• Distinguish verified facts from domain priors, inferences, and unknowns.
• Investigate every material REVISE or BLOCK finding. Fix it or establish contrary evidence, then rerun the relevant review.
• Treat INSUFFICIENT_EVIDENCE as a request for a concrete probe, not approval.
• Keep edits and implementation in Codex. Opus must return all useful guidance in its response.

For a completion gate, start a fresh call after implementation and tests. Include the final diff scope, verification commands and results, and known limitations. An APPROVE verdict is advisory input; Codex still needs independent evidence that the requested goal is satisfied.

If the launcher cannot run, report that the advisor was unavailable. Do not fabricate an advisory verdict.