advisor-reviewlisted

Run a runtime-selected advisor-style review in the current task when the user invokes $advisor-review, says Advisor Review or @Advisor-Review, explicitly asks for an advisor/second opinion/adversarial review, or task instructions explicitly require independent review. Uses source-anchored evidence, bounded artifacts, and an adopt/reject/defer decision gate; do not substitute tool discovery, another task, or same-agent reflection.
mrcha033/skills · ★ 0 · AI & Automation · score 57

Install: claude install-skill mrcha033/skills

# Advisor Review Use the bundled scripts from this skill directory. An explicit invocation is a latch: run the reviewer in the **same parent task** before a substantive final answer or an action that follows or conflicts with the review. Read-only evidence collection may continue while constructing the packet. Do not search the global tool catalog for an advisor. Do not open a second user-visible task. The only child execution is the isolated runtime CLI process started by `scripts/run_advisor.py`. ## Requirements Require shell execution, an authenticated CLI for the active session runtime, access to the caller-selected advisor model, and outbound access for that runtime from the parent shell. Supported runtimes are `codex`, `claude`, `opencode`, and `gemini`; `--runtime auto` uses the active session marker such as `OPENCODE=1`, and refuses an ambiguous host instead of guessing. A parent sandbox that blocks subprocess networking also blocks the nested reviewer. If any requirement is missing, report that the independent review is blocked; do not weaken the sandbox without user or task authority. The runner is not a confidentiality boundary. Sanitize evidence and artifacts before building the packet. ## Workflow 1. Announce that Advisor Review is being used and why. 2. Select one phase: - `plan`: challenge a proposed approach. - `stuck`: diagnose failed or looping work. - `pivot`: compare a materially different approach. - `final`: audit completed work and it