ffroliva/gflow-cli/skills/llm-council/SKILL.md
Use when a pr-council-review (PR or branch mode) result needs independent corroboration from a different model family before trusting a GREEN verdict — high-stakes, security-sensitive, or architecturally significant reviews where same-model-family Claude subagents might share a blind spot. Also use when the user asks for "external", "second opinion", "cross-model", or names codex/antigravity (`agy`) or another external CLI coding agent alongside a review.
Decision brief
Also use when the user asks for "external", "second opinion", "cross-model", or names codex/antigravity (`agy`) or another external CLI coding agent alongside a review.
Compatibility matrix
| Platform | Status | Evidence | What to check |
|---|---|---|---|
| Codex | Declared | Source record | Install path and trigger |
| Claude Code | Not declared | No explicit evidence | Portability before use |
| Cursor | Not declared | No explicit evidence | Portability before use |
| Gemini CLI | Not declared | No explicit evidence | Portability before use |
Installation
The source command is displayed only when detected. A safe inspection prompt is always available so your agent can explain every action before execution.
npx skills add https://github.com/ffroliva/gflow-cli --skill "skills/llm-council"Inspect the Agent Skill "llm-council" from https://github.com/ffroliva/gflow-cli/blob/c64a8d8b246f9551e8a96c7fcf363d15e045c9b2/skills/llm-council/SKILL.md at commit c64a8d8b246f9551e8a96c7fcf363d15e045c9b2. List every install step, command, network request, credential, file read/write, external action, and rollback step. Explain whether it fits my task. Do not install or execute anything until I approve.
Workflow
Any pr-council-review run (PR mode or branch mode) where the artifact is high-stakes enough to want a second, differently-biased opinion before calling it GREEN.
Tier controls tool breadth, not review rounds. Fix → re-verify → repeat until GREEN (or a round cap) happens at every tier — that's how council review works, not a tier knob.
Fixed, tested invocation recipes. Do not improvise a command for a listed tool — the "obvious" invocation is often a trap (see codex below).
NEVER codex review. Its built-in prompt has gotten stuck in a self-inflicted loop reading skill files via a malformed PowerShell command, then retrying the identical broken command for 20+ minutes with zero progress. Co…
What it is: Google's Antigravity harness, invoked via agy. It supplies the high-tier's second, different-model-family opinion.
Permission review
The documentation asks the agent to read local files, directories, or repositories.
**Use:** `codex exec -s read-only -C <absolute-repo-dir> --skip-git-repo-check "<direct, fully self-contained prompt>"`.The documentation asks the agent to read local files, directories, or repositories.
| Running `codex review` because it sounds like the obvious subcommand for a review task | Use `codex exec -s read-only -C <dir> --skip-git-repo-check "<prompt>"` — see registry |Evidence record
| Signal | Value | Evidence type | Meaning |
|---|---|---|---|
| Quality score | 84/100 | Computed | Documentation, specificity, maintenance, and trust rules |
| Repository stars | 96 | Source | Repository attention, not individual Skill quality |
| Compatibility | 1 platforms | Source | Declared in the catalog source record |
| Usage guide | automated source guide | Editorial | Generated or reviewed according to the visible evidence level |
Pinned source
llm-council — external-tools review layerWraps pr-council-review (unchanged) and adds a layer of external CLI coding agents (codex, plus Antigravity — the agy harness) as additional independent reviewers, then folds their verdicts into the same synthesis. Internal Claude subagents are independent per-dimension but share one model family's blind spots — a phrasing choice, a Windows-vs-POSIX nuance, or a syntax error that reads fine to one Claude reviewer reads fine to all of them. A different model family catches a different error distribution. Confirmed live: on one review, external tools caught 6 real, distinct issues (a wording-accuracy bug, a wrong test count, a Windows-only test-triviality nuance, a test-isolation gap, a missing test, a malformed markdown fence) that 12 internal Claude-subagent dispatches (6 dimensions × 2 rounds) had all missed.
pr-council-review run (PR# mode or branch mode) where the artifact is high-stakes enough to want a second, differently-biased opinion before calling it GREEN./review (single-agent) for that; pr-council-review alone for a normal-stakes PR.| Tier | Internal (pr-council-review) | External tools |
|---|---|---|
small (default) | ✅ full dimension council | none — identical to running pr-council-review directly |
medium | ✅ | codex |
high | ✅ | codex + Antigravity (agy) |
Tier controls tool breadth, not review rounds. Fix → re-verify → repeat until GREEN (or a round cap) happens at every tier — that's how council review works, not a tier knob.
Fixed, tested invocation recipes. Do not improvise a command for a listed tool — the "obvious" invocation is often a trap (see codex below).
codexcodex review. Its built-in prompt has gotten stuck in a self-inflicted loop reading skill files via a malformed PowerShell command, then retrying the identical broken command for 20+ minutes with zero progress. Confirmed reproducible on a clean retry.codex exec -s read-only -C <absolute-repo-dir> --skip-git-repo-check "<direct, fully self-contained prompt>".codex --version (near-instant; confirms binary health only, not auth/quota).xhigh) reasoning effort. Always background it — never block synchronously.codex exec can leave codex.exe / codex-code-mode-host.exe / sandbox-helper processes running on Windows. After any kill, verify via tasklist/ps that the named PIDs are actually gone before retrying — a retry racing an orphan still writing the same output path silently corrupts the result.agy)agy. It supplies the high-tier's second, different-model-family opinion.agy --agent <gsd-agent-name> --new-project --add-dir <absolute-repo-dir> -p "<prompt>" (run agy agents to list available agent names, e.g. gsd-plan-checker for reviewing a plan).agy --version first regardless — a fast binary-health check.--add-dir; a bare unexplained termination error even with --agent/--new-project) — plausibly account-quota exhaustion, unconfirmed. Don't silently retry past its probe. When agy isn't installed or won't run non-interactively, suggest installing it (or substituting another external CLI coding agent — e.g. a codex-only medium run) and continue best-effort with whatever did return; never block the whole round on it.pr-council-review (unchanged) for the internal dimension council.pr-council-review produces — same GREEN/YELLOW/RED vocabulary, tagged by source (e.g. D3 (internal) vs codex (external)). A tool that fails or times out after passing its probe is dropped from this round with an explicit note in the report. Never silently drop, never block the whole round on one flaky tool.pr-council-review's existing shape, plus an "External tools" line noting which ran / were skipped / failed and why.| Mistake | Fix |
|---|---|
Running codex review because it sounds like the obvious subcommand for a review task | Use codex exec -s read-only -C <dir> --skip-git-repo-check "<prompt>" — see registry |
| Dispatching the real (slow) external call before probing | Probe first, short timeout — a dead tool costs 10-20 min discovered late vs. ~15s discovered early |
| Treating one failed external tool as a reason to abandon the whole external layer | Best-effort: drop that tool for this round, disclose it, keep going with whatever did return |
| Blocking synchronously on an external tool call | Always background it — internal dimensions and other external tools shouldn't wait |
| Retrying a timed-out tool without checking for orphaned processes first | tasklist/ps check + explicit kill before any retry against the same output path |
Silently downgrading to codex-only when agy is unavailable | Disclose the drop and suggest installing Antigravity or substituting another external CLI agent — don't hide the reduced coverage |
REQUIRED SUB-SKILL: the internal council dispatch is pr-council-review (skills/pr-council-review/SKILL.md) — this skill does not reimplement dimension detection, synthesis rules, or report shape, it wraps them.
Alternatives
xiaolai/nlpm
Universal NL programming conventions — SKILL.md open spec (agentskills.io), AGENTS.md as canonical universal memory file, vague-quantifier list, prompt engineering layers, naming conventions, the override system. Tool-specific schemas live in nlpm:conventions-claude / nlpm:conventions-codex / nlpm:conventions-antigravity.
PramodDutta/qaskills
Generate optimized test combinations using pairwise (all-pairs) testing algorithms to achieve maximum coverage with minimum test cases across multiple input parameters
PramodDutta/qaskills
Gate RAG pipelines in CI with versioned golden eval sets, per-metric thresholds, baseline drift detection, and a build that fails when retrieval or answer quality regresses.
PramodDutta/qaskills
Generate comprehensive test cases from state machine models covering all states, transitions, guard conditions, and invalid transition attempts for workflow-heavy features