--- name: counselors description: Run multi-agent code reviews with the counselors CLI and synthesize actionable findings. compatibility: opencode --- # Counselors Skill Use this skill when the user asks for second opinions, multi-model review, architecture critique, or bug/risk hunting. ## Core workflow 1. Gather context quickly. - Identify the target files and question scope. - If needed, inspect `git diff HEAD` and `git diff --staged`. 2. Discover configured tools and groups. - Run: - `counselors ls` - `counselors groups ls` - Show the full outputs to the user and ask which tools/group to use. 3. Build prompt file with counselors. - Use `counselors mkdir --json` and pipe prompt content in. - Use `@path/to/file` references instead of inlining large files. 4. Dispatch run. - Single pass: `counselors run -f --tools --read-only best-effort --json` - Iterative pass: `counselors loop -f --tools --read-only best-effort --json` - Preset loop: `counselors loop --preset "" --tools --read-only best-effort --json` 5. Read outputs and synthesize. - Parse the JSON manifest. - Read each `outputFile` for successful tools. - Check `stderrFile` for failed/empty tools. - Produce concise synthesis: - consensus - disagreements - key risks - recommendation - Include the saved output directory path. ## Usage guardrails - Ask for tool/group selection before dispatch. - Confirm final tool list before running. - Prefer `--json` output for parseability. - Use long timeouts for dispatch (often 10+ minutes). - If selected tools include custom `opencode-*`, avoid `--read-only strict` because those tools are configured as `bestEffort` and will be skipped under strict policy. - If no tools are configured, instruct user to run `counselors init --auto`. ## Useful commands - `counselors init --auto` - `counselors tools add` - `counselors ls` - `counselors groups ls` - `counselors run --help` - `counselors loop --help`