agents-consilium

name

agents-consilium

description

Query external AI agents (Codex, Gemini, OpenCode, Claude Code headless) in parallel for independent second opinions, code review, bug investigation, and consensus on high-stakes decisions. Agents and models are configurable in config.json. Use for architecture choices, security review, or ambiguous problems where independent perspectives matter. Not for simple questions answerable from docs or the codebase — use web search or repo exploration instead.

Consilium: Multi-Agent Orchestration

Query external AI agents for independent, unbiased expert opinions. Each agent has a distinct thinking role and responds in a structured format for easy comparison.

Why this skill

Different frontier models see different things. Each has a slightly different training distribution, tool-use style, and failure mode — so they latch onto different aspects of the same problem.

• Brainstorming / problem-solving / feature design. Querying Codex + Claude + OpenCode/Gemini (or any subset) in parallel yields a wider solution space than any single model alone. You get original, non-obvious alternatives that one model would never surface on its own.
• Code review. Different models find different issues. One catches a subtle race condition; another flags an auth gap; a third questions the architecture. The union of their findings is materially broader than a single-reviewer pass.

The skill keeps each agent independent (no debate, no cross-contamination) and lets the caller adjudicate — you get raw parallel perspectives, not a homogenized committee answer.

Contents

• Quick Start
• Shell escaping — read this before passing prompts inline
• Design Principles
• Anti-Bias Protocol
• Agent Freedom and Read-Only Guardrails
• Configuration
  • OpenCode provider choice
  • Claude Code backend
• Scripts
  • Flags & Exit Codes
• Code Review Mode
• Multi-Stage Review Modes: superreview & ultrareview
• When to Use Which
• Synthesizing Responses
• Prompt Patterns
• Environment Variables
• Prerequisites

Quick Start

# 1. See what's configured (XML plan — dry-run, no agents run).
scripts/consensus-query.sh --list-agents

# 2. Ask the consensus (human-readable markdown).
scripts/consensus-query.sh "Should we use Postgres or SQLite for this CLI tool?"

# 3. Agent-friendly output (stable XML, escaped via CDATA).
scripts/consensus-query.sh --xml "Review this function" < src/auth.py

# 4. Code review mode (2 specialists, quoted-code validated, XML or markdown).
scripts/code-review.sh path/to/file.py
git diff HEAD | scripts/code-review.sh --xml --diff

Edit config.json to enable/disable agents or swap models. See config.example.json for a fuller template with multiple backends.

Passing the prompt

The prompt is a positional argument. Three ways to pass it — prefer file-based forms for anything containing backticks, $, !, or quotes (see Shell escaping):

# (a) RECOMMENDED: From a file via flag — RAW mode
#     (no role/principles/template wrapping; agents see the file verbatim).
#     Safe for any prompt content. Use this for benchmarks/evals too.
scripts/consensus-query.sh --xml --prompt-file prompt.txt

# (b) RECOMMENDED: From a file via stdin — best for long multi-line prompts.
#     With NO positional argument, stdin is treated as the prompt.
#     Template wrapping (principles + role + output schema) is applied.
scripts/consensus-query.sh --xml < prompt.txt
cat prompt.txt | scripts/consensus-query.sh --xml

# (c) Inline string — ONLY for short, shell-safe prompts
#     (no backticks, no $, no !, no embedded quotes).
scripts/consensus-query.sh --xml "review this design"

When BOTH a positional prompt and stdin are given, stdin is appended to the prompt as --- Input --- context. That is the existing pattern for piping a file under review:

cat src/auth.py | scripts/consensus-query.sh "review this code"
#       └── stdin = context ────┘   └── positional = the prompt ─┘

Shell escaping — read this before passing prompts inline

The trap. Technical prompts routinely contain backticks (around config keys, function names, code paths), $ (env vars, regex anchors, $() substitution), ! (interactive history), and unbalanced quotes. Inside a double-quoted positional argument, bash/zsh will execute the contents of backticks and $(...) as commands, expand $variable, and silently splice the result back into the prompt. Failed substitutions become empty strings. The agent then receives a mangled prompt and — worse — codex exec (and claude -p, opencode run) may stall in the foreground waiting on stdin for clarification, holding the process at 0% CPU for hours before you notice.

The exact symptom in ~/.codex/log/codex-tui.log or the agent's transcript:

(eval):1: command not found: index.knn.advanced.approximate_threshold=-1
(eval):1: command not found: POST
Reading additional input from stdin...

This is a shell-quoting bug, not a CLI bug. Single quotes around EOF in a heredoc, or a separate file, are the fix.

DO / DON'T

Don't	Do
tool "What does `foo()` do?" — backticks get eval'd by the shell	Use one of the three safe patterns below
tool "Explain $PATH precedence" — $PATH expands to your env var	Same
tool "Why does $(date) appear?" — $(date) runs as a subshell	Same
tool "Don't run !! again" — !! triggers history (zsh interactive)	Same
Inline prompts with any of: `, $, $(, !, embedded "	Files or heredocs

Three safe patterns (in order of preference)

# (1) RAW mode via --prompt-file — safest. No shell interpretation at any layer.
#     No template/role wrapping either; agents see the file verbatim.
scripts/consensus-query.sh --xml --prompt-file prompt.md

# (2) Pipe via stdin — template wrapping (principles + role + output schema) IS applied.
cat prompt.md | scripts/consensus-query.sh --xml
scripts/consensus-query.sh --xml < prompt.md

# (3) Heredoc with SINGLE-quoted delimiter — single quotes around EOF disable
#     expansion of $, backticks, !, and \ inside the body. This is the
#     canonical bash idiom for "embed an arbitrary technical prompt inline".
scripts/consensus-query.sh --xml "$(cat <<'EOF'
Explain how `index.knn.advanced.approximate_threshold=-1` interacts with
`POST /_forcemerge?max_num_segments=1` in OpenSearch 3.5. Cover $variable
expansion semantics in the config parser too.
EOF
)"
#                ^^^^                                            ^^^^
#                single quotes around EOF — REQUIRED for safety

If you write <<EOF (no quotes) or <<"EOF", double-quoted-style expansion happens INSIDE the heredoc. Always single-quote the opening delimiter when the body is a prompt.

Bypassing the skill (calling codex exec / claude -p directly)

The same trap applies the moment you build a codex exec -m … "<prompt>" or claude -p "<prompt>" invocation yourself — e.g. because you want a non-default model, a custom -C <dir>, or a one-off -o <transcript>. Wrap the prompt the same way:

# Safe: prompt loaded from a file
codex exec -m gpt-5.6-sol --full-auto -C /tmp/work -o "$OUT" "$(cat prompt.md)"

# Safe: single-quoted heredoc
codex exec -m gpt-5.6-sol --full-auto -C /tmp/work -o "$OUT" "$(cat <<'EOF'
…prompt body with `backticks` and $vars…
EOF
)"

# UNSAFE: inline string with shell-special characters
codex exec -m gpt-5.6-sol --full-auto -C /tmp/work -o "$OUT" \
  "Explain \`index.knn.advanced.approximate_threshold=-1\` semantics"
#          ^^ even escaping the backticks here is fragile —
#             the shell still has to interpret them once

Diagnostic recipe

If codex / claude / opencode appears to hang at 0% CPU and the output file is empty (or ~146 bytes of garbage prelude), this is almost always the bug. Check, in order:

1. tail -50 ~/.codex/log/codex-tui.log (or the agent's transcript / your captured 2> stderr) for command not found: or Reading additional input from stdin....
2. The prompt as the shell saw it: re-run the same command with set -x (or prepend echo to a copy) to print the expanded argv. If a backtick'd token vanished, expansion ate it.
3. Kill the stuck process (kill -TERM <pid>; kill -KILL if needed), reissue the prompt via --prompt-file or a single-quoted heredoc.

Design Principles

Intellectual independence: Agents are instructed to think from first principles, challenge the framing of questions, and propose alternatives not mentioned in the query. They are free thinkers within the given context, not yes-men.

Role differentiation (set per agent in config.json):

• analyst = Rigorous Analyst — precision, code correctness, edge cases, implementation depth, security (default for Codex)
• lateral = Lateral Thinker — cross-domain patterns, creative alternatives, questioning premises, big picture (default for Gemini / OpenCode with Gemini-3.1-Pro)

Structured output: All agents respond using a common template (Assessment, Key Findings, Blind Spots, Alternatives, Recommendation with confidence level), making synthesis straightforward.

Anti-Bias Protocol

When formulating queries for consilium, follow these rules to maximize the value of independent opinions:

1. State the problem, not your solution. Instead of "Should we use X?", describe the constraints and goals.
2. Don't lead. Avoid "I think X is best, what do you think?" — this anchors the response.
3. Include raw context. Pipe code files or paste error logs directly rather than summarizing them (summaries carry your interpretation).
4. Omit your hypothesis when possible. Let agents form their own before revealing yours.

Agent Freedom and Read-Only Guardrails

Agents are spawned in the caller's current working directory with their native agentic toolchain intact. They can:

• Read, Grep, Glob, find_references, git log/blame across the real repository
• Consult CLAUDE.md, AGENTS.md, README, config files, tests, call sites, neighboring modules
• Use web search / fetch if their backend supports it (Claude Code, OpenCode, Codex all do)
• Run SAST-style introspection via their built-in shells

What they cannot do (enforced per backend):

Backend	Read-only guard
Codex	--sandbox read-only --ask-for-approval never
Claude Code	--permission-mode plan
OpenCode	--agent plan (plan is opencode's built-in read-only agent)
Gemini CLI	--approval-mode plan

No Edit, Write, Bash(git commit ...), Bash(rm ...), or any write-back tool is authorized. Implementation of recommendations is the caller's job. If a backend tries to escalate (e.g. needs to run a command that violates read-only), the call fails rather than silently escalating.

Configuration

Agents are declared in config.json at the skill root. Each agent has:

Field	Purpose
enabled	Whether it participates in consensus-query
backend	CLI that actually runs: codex-cli, gemini-cli, opencode, claude-code
model	Model id passed to that CLI
role	analyst (deep/precise) or lateral (broad/creative)
label	Display name in reports (optional)
effort	Reasoning effort. codex-cli: maps to model_reasoning_effort (minimal, low, medium, high, xhigh). opencode: maps to opencode run --variant — provider-specific, see Discovering reasoning variants. claude-code: maps to claude --effort (low, medium, high, xhigh, max). Gemini ignores it.

Default config (config.json):

• codex (backend=codex-cli, model=gpt-5.6-sol, effort=high, role=analyst) — enabled and the default code-review agent
• codex-gpt-5.6 (backend=codex-cli, model=gpt-5.6, effort=high, role=analyst) — disabled; latest alias routing to Sol
• codex-gpt-5.6-terra (backend=codex-cli, model=gpt-5.6-terra, effort=medium, role=analyst) — disabled
• codex-gpt-5.6-luna (backend=codex-cli, model=gpt-5.6-luna, effort=low, role=analyst) — disabled
• gemini-cli (backend=gemini-cli, model=gemini-3.1-pro-preview, role=lateral) — disabled
• opencode (backend=opencode, model=opencode-go/glm-5.2, role=lateral, effort=max) — enabled and the default OpenCode agent
• claude-code (backend=claude-code, model=claude-sonnet-5, effort=high, role=analyst) — disabled
• claude-sonnet (backend=claude-code, model=claude-sonnet-5, effort=high, role=lateral) — disabled
• opencode-go-minimax (backend=opencode, model=opencode-go/minimax-m3, role=lateral, effort=thinking) — enabled
• opencode-go-kimi (backend=opencode, model=opencode-go/kimi-k2.7-code, role=analyst, effort=none) — enabled
• opencode-go-glm (backend=opencode, model=opencode-go/glm-5.1, role=lateral, effort=none) — enabled
• opencode-go-qwen37-max (backend=opencode, model=opencode-go/qwen3.7-max, role=analyst, effort=none) — enabled
• opencode-go-qwen37-plus (backend=opencode, model=opencode-go/qwen3.7-plus, role=lateral, effort=none) — enabled

Effort policy: use the highest supported OpenCode variant for models that expose one (opencode-go/glm-5.2 uses max, opencode-go/minimax-m3 uses thinking). Use none for models whose variants list is empty (glm-5.1, kimi-k2.7-code, qwen3.7-max, qwen3.7-plus) so the backend omits --variant entirely.

Codex GPT-5.6 model profiles

Model	Intended use	Default profile
gpt-5.6	Latest GPT-5.6 alias; the wrapper resolves it to Sol for Codex CLI compatibility	codex-gpt-5.6, high, disabled
gpt-5.6-sol	Flagship tier for the hardest quality-first coding and reasoning workflows	codex, high, enabled
gpt-5.6-terra	Balanced cost, latency, and quality	codex-gpt-5.6-terra, medium, disabled
gpt-5.6-luna	High-throughput, simple, or strict-latency tasks	codex-gpt-5.6-luna, low, disabled

code-review.sh assigns enabled agents in config order. Keeping codex first makes gpt-5.6-sol at high effort the default security reviewer while preserving the stable codex agent id used by scripts and existing callers. Select another profile ad hoc with -a codex-gpt-5.6-terra or -a codex-gpt-5.6-luna.

The Codex CLI wrapper normalizes gpt-5.6 to gpt-5.6-sol before execution. This preserves the documented alias semantics on ChatGPT-backed Codex installations whose server-side alias roster has not caught up yet.

Multiple agents can share one backend — the dispatcher passes the entry id through CONSILIUM_AGENT_ID, so each backend script reads its own slice of config.json.

Edit config.json to flip agents on/off or change models. Set CONSILIUM_CONFIG=/path/to/custom.json to use an override file.

OpenCode provider choice

The opencode backend works with any provider/model that OpenCode supports. The default configuration uses the OC-Go provider:

• OC-Go (default): "model": "opencode-go/glm-5.2" — goes through OpenCode's OC-Go provider. Works once OpenCode is authenticated for OC-Go.

Flip between providers by editing the model field; the rest of the config stays the same.

Discovering OpenCode reasoning variants per model

opencode run --variant <effort> is provider-specific — each model exposes its own set (or none). Don't guess: enumerate them from the CLI before setting effort in config.json.

One-liner — list every model with its supported variants:

opencode models opencode --verbose 2>&1 | python3 -c '
import sys, json
lines = sys.stdin.read().split("\n")
i = 0
while i < len(lines):
    line = lines[i].strip()
    if line.startswith("opencode/") or line.startswith("opencode-go/"):
        model_id, json_lines, depth, started = line, [], 0, False
        i += 1
        while i < len(lines):
            s = lines[i]; json_lines.append(s)
            for c in s:
                if c == "{": depth += 1; started = True
                elif c == "}": depth -= 1
            i += 1
            if started and depth == 0: break
        try:
            v = list(json.loads("\n".join(json_lines)).get("variants", {}).keys())
            print(f"{model_id}\t{v}")
        except Exception: pass
    else:
        i += 1
'

Swap opencode for opencode-go (or any other provider id) to scan a different namespace; drop the provider arg to scan everything opencode models knows.

Interpreting the result:

• Non-empty list (e.g. ['low', 'medium', 'high', 'max']) → set effort to the highest one you want.
• [] → the model has no reasoning variants. --variant is silently ignored; setting effort in config is harmless but does nothing.
• If a variant in your config isn't on the list, opencode run rejects the call. Re-enumerate after upgrading opencode — providers add/remove tiers between releases.

Snapshot of the currently configured opencode models (re-run the one-liner if you change the set):

Model	Variants	effort in default config
opencode-go/glm-5.2	high, max	max
opencode-go/glm-5.1	—	none
opencode-go/kimi-k2.7-code	—	none
opencode-go/minimax-m3	none, thinking	thinking
opencode-go/qwen3.7-max	—	none
opencode-go/qwen3.7-plus	—	none

Claude Code backend

The claude-code backend shells out to claude -p (headless mode, see docs). Useful when you want a second Claude in the consilium — e.g. Sonnet 5 as analyst cross-checking Codex.

• model: a shortname (sonnet, haiku) or full id (claude-sonnet-5).
• effort: maps to claude --effort — accepts low, medium, high, xhigh, max. Default config sets high for Sonnet; omit the field to use the backend fallback of high, which also applies when selecting Opus through CLAUDE_MODEL.
• Runs in the caller's CWD with --permission-mode plan — Claude can freely Read/Grep/Glob/Bash read-only across the project, but cannot Edit/Write. Override with CLAUDE_PERMISSION_MODE only if you know what you're doing.
• Authentication uses the same Claude Code credentials the CLI is already logged in with (claude /login).

Note: claude-code is disabled in the default config to avoid spawning another Claude session accidentally. Flip enabled to true in config.json (or CONSILIUM_CONFIG) when you want it in the consensus run.

Scripts

All scripts in scripts/ directory. The skill auto-detects its install location.

Single Agent Queries

Per-agent scripts always execute when invoked. The enabled field in config.json is consulted only by consensus-query.sh to build the default agent set (when neither -a nor -x is given). Direct invocation of a per-agent script ignores enabled — that's by design (single source of truth for the run/skip decision lives in the dispatcher).

When -a/-x causes consensus-query.sh to run an enabled=false agent, the dispatcher emits a stderr line like [<Label>] forced via --agents (enabled=false in config) so the override is visible.

# Codex (analyst by default)
scripts/codex-query.sh "question" [context_file]
cat file.py | scripts/codex-query.sh "review this"

# Gemini CLI (lateral by default; disabled in default config)
scripts/gemini-query.sh "question" [context_file]
cat file.py | scripts/gemini-query.sh "review this"

# OpenCode (lateral by default, model per config.json)
scripts/opencode-query.sh "question" [context_file]
cat file.py | scripts/opencode-query.sh "review this"

# Claude Code (analyst by default; disabled in default config)
scripts/claude-query.sh "question" [context_file]
cat file.py | scripts/claude-query.sh "review this"

Consensus Query (All Enabled Agents in Parallel)

scripts/consensus-query.sh "architecture question"
cat file.py | scripts/consensus-query.sh "review this code"
scripts/consensus-query.sh --xml "review this"            # XML report for agent consumers
scripts/consensus-query.sh --list-agents                   # dry-run: dump plan, don't query

consensus-query.sh reads config.json, launches every agent with enabled=true in parallel, and prints their responses grouped by label. Add/remove agents permanently by editing the config; for ad-hoc runs use -a/--agents and -x/--exclude (see below).

Flags & Exit Codes

All scripts accept -h / --help. Both consensus-query.sh and code-review.sh accept:

Flag	Effect
--xml	Emit <consilium-report> (or <code-review-report>) with each agent wrapped in <agent>…<response><![CDATA[…]]></response></agent>. Stable for agent consumers (no markdown-heading collision).
--list-agents (consensus only)	Print <consilium-plan> (every configured agent, enabled/disabled, with backend-available) and exit. No queries are run — use this as an inspection / dry-run.
-a, --agents <ID|GLOB>	Override the active agent set with this id or glob (e.g. 'opencode-go-*'). Repeatable; comma-separated values also accepted (-a codex,opencode-go-kimi). When given, the per-agent enabled flag in config.json is ignored — only matched agents run. Falls back to env CONSILIUM_AGENTS.
-x, --exclude <ID|GLOB>	Subtract matching agents from the active set. Repeatable. Combine with --agents for include-then-exclude composition. Falls back to env CONSILIUM_EXCLUDE.

Ad-hoc agent selection examples:

# Single agent
scripts/consensus-query.sh -a opencode-go-kimi "Q"

# All OC-Go models (glob)
scripts/consensus-query.sh -a 'opencode-go-*' "Q"

# Everything-except-codex
scripts/consensus-query.sh -x codex "Q"

# Composition: only OC-Go but skip MiniMax
scripts/consensus-query.sh -a 'opencode-go-*' -x opencode-go-minimax "Q"

# Same via env (scriptable)
CONSILIUM_AGENTS='codex,opencode-go-kimi' scripts/consensus-query.sh "Q"

Exit codes (stable across all scripts):

Code	Meaning
0	Success (all queried agents replied; or, for consensus-query.sh, the active agent set may be smaller than the configured set if some are disabled or filtered)
2	Consensus only: partial failure (≥1 succeeded, ≥1 failed)
3	Consensus only: every queried agent failed
4	Config error (missing CLI, invalid config, unknown role/agent id)
5	Usage error (missing prompt, unknown flag)
other	Propagated from the backend CLI (e.g. 124 on timeout)

Code Review Mode

scripts/code-review.sh is a focused pipeline for reviewing a single file or a unified diff. It runs exactly two specialist passes — security and correctness — in parallel, then validates each finding's quoted-code against the real source.

Design choices are grounded in the 2024-2026 multi-agent code review literature:

• Two specializations only (security + correctness). Readability/perf agents empirically produce nit spam and hurt precision.
• No coordinator / no debate. The caller (you) adjudicates. Debate rounds empirically entrench errors (Wu et al. 2025; Choi et al. 2025).
• Heterogeneous models via the existing config (Codex + OpenCode by default) reduce shared blind spots.
• Fixed cost. Adding a 3rd enabled agent does not add a 3rd pass; the skill always runs 2 passes and rotates agents round-robin.
• Hallucinated line numbers are caught locally. Every finding carries <quoted-code>, and the validator cross-checks it against the source file (quote-valid="true|false").

Usage

# File on disk (quoted-code validated against the file)
scripts/code-review.sh path/to/file.py
scripts/code-review.sh --xml path/to/file.py

# Unified diff piped on stdin (quoted-code validation is skipped)
git diff HEAD | scripts/code-review.sh --diff
git diff HEAD | scripts/code-review.sh --xml --diff

Finding schema (XML output)

<finding index="N" severity="critical|high|medium|low" category="security|correctness"
         file="..." line-start="N" line-end="N" confidence="0.0..1.0"
         source-agent="..." source-role="security|correctness"
         quote-valid="true|false">
  <title>...</title>
  <rationale><![CDATA[includes one reason this might be a false positive]]></rationale>
  <suggested-fix><![CDATA[...]]></suggested-fix>
  <quoted-code><![CDATA[verbatim source at line-start..line-end]]></quoted-code>
</finding>

Findings are sorted severity desc, confidence desc. No severity filtering by default — triage is the caller's job.

Severity rubric

Unified across security + correctness. Specialists score each finding on two axes (worst-case impact × likelihood/reachability) and pick the tier that matches. Synthesized from CVSS v4, OWASP Risk Rating, GitHub Advisory DB, Chromium, MSRC, SEI CERT, SonarQube, Semgrep.

Severity	Action horizon	Operational definition	Security examples	Correctness examples
critical	Merge blocker	RCE / trust-boundary bypass / data loss / guaranteed outage, with a concrete exploit or dataflow trace	SQLi on public endpoint with concatenated query; unsafe deserialization of untrusted input; hardcoded prod credential	Payment/ledger math silently corrupts balances; unconditional null deref on hot request path; race on shared mutable state under prod load
high	Fix before release	Critical-tier impact gated by a non-trivial precondition (auth, specific config), OR moderate impact with high reachability	Stored XSS in authenticated admin view; CSRF on state-changing endpoint; path traversal behind login; missing authz on tenant resource	Unhandled exception on documented error path crashing a worker; file/DB-handle leak exhausting pools; retry logic that double-charges
medium	Schedule	Limited impact (info disclosure, localized incorrectness, degraded-but-recoverable), OR critical impact gated by implausible preconditions	Stack traces leaked to end users; missing HttpOnly/Secure on non-session cookie; weak-but-not-broken crypto parameter	Incorrect edge-case handling in non-critical helper; missing input validation that callers already satisfy; N+1 query degrading a list endpoint
low	Optional / backlog	Cosmetic, stylistic, defense-in-depth; minimal real-world impact	Missing nosniff header where CSP already mitigates; Math.random() for non-security id	Dead code; inconsistent naming; redundant null check after non-null assertion

Adjustments: downgrade one level on mitigating factors (auth required, non-default config, unusual interaction). Speculative findings stay at the lower tier — upgrade only with a working PoC or trace.

Using the results (for the caller)

You are the adjudicator. Specialists emit independent findings — your job is to select and synthesize, not re-review (RovoDev 2601.01129, RevAgent 2511.00517).

1. Drop quote-mismatched findings (quote-valid="false") — likely hallucinations.
2. Merge duplicates across specialists. Same root cause in different framings → one item; keep the clearer rationale and note both agents.
3. Surface conflicts without resolving them. If Security says "sanitize X" and Correctness says "X is fine" — present both to the user and let them adjudicate; don't break the tie yourself.
4. Gate by action horizon using the severity rubric above: critical = block the merge, high = fix before release, medium = track, low = optional.
5. Do not re-review. Do not generate new findings inside the adjudication step. Do not run a debate loop — adversarial re-reviewing empirically reduces precision (CR-Bench 2603.11078).

When NOT to use code-review mode

• Open-ended architecture questions → use consensus-query.sh; specialists will be too narrow.
• Huge files (>1000 lines) → split into function-sized diffs first; LLMs degrade past that length.
• Multi-file cross-references → not modelled here; rerun per file and stitch findings.

Multi-Stage Review Modes: superreview & ultrareview

code-review.sh is single-stage and caller-judged. For higher-stakes reviews where you want the union of many panels filtered automatically by an LLM judge, the skill ships two multi-stage pipelines ported from the ultrareview-bench (see docs/blog/code-review-2pass-pilot/ if you have access). Each one prescribes a fixed agent set and stage layout — they're not configurable per-call, by design, because the configurations were tuned by marginal-uplift analysis on a 65-issue ground-truth pilot.

Important: these are heavy modes. Don't run them on every diff. Use them when you'd otherwise pull two senior engineers off other work for a deep review, or for code that touches money / auth / persistence.

scripts/superreview.sh — small-swarm + 2 frontier add-ons

10 LLM calls; ~$0.90–1.50 on a 12KB file (linear with size). Pareto sweet-spot in the bench at 67.7% recall / 82.7% sev-w on snippet1.cs.

Stage 1: discovery-small (parallel)    7 OC-Go passes
  - opencode-go-minimax      analyst       uncapped
  - opencode-go-qwen37-plus  analyst       uncapped
  - opencode-go-qwen37-plus  lateral       uncapped
  - opencode-go-glm          architecture  uncapped
  - opencode-go-glm          correctness   cap=10
  - opencode-go-qwen37-max   architecture  cap=3
  - opencode-go-qwen37-max   security      uncapped
Stage 2: discovery-frontier (parallel) 2 hand-picked add-ons
  - opencode                   analyst       uncapped
  - claude-code (Sonnet 5 high) lateral      uncapped
Stage 3: dedup (deterministic union)
Stage 4: judge — claude-sonnet (default)

Usage:

scripts/superreview.sh path/to/file.cs
scripts/superreview.sh --xml path/to/file.cs
git diff HEAD | scripts/superreview.sh --diff
scripts/superreview.sh --dry-run path/to/file.cs   # plan + config check, no LLM calls
scripts/superreview.sh --judge claude-code path/to/file.cs   # override judge

scripts/ultrareview.sh — broad-grid + specialists + probe

21 LLM calls; ~$1.50–3.00 on a 12KB file. Best severity-weighted recall in the bench (86.4%). Slower and more expensive than superreview; use when you need maximum coverage and lowest false-positive rate.

Stage 1: broad (parallel)         4 frontier analysts
  - codex (gpt-5.6-sol high)     analyst      uncapped
  - claude-code (Sonnet 5 high)  analyst      uncapped
  - opencode (GLM-5.2)           lateral      uncapped
  - opencode-go-qwen37-max       analyst      uncapped
Stage 2: specialists (parallel)   5x3 matrix, uniform cap=10
  - 3 small models × 5 roles (security/correctness/performance/architecture/consistency)
Stage 3: probe (sequential)       1 generic gap probe (model picks focus)
  - opencode-go-glm auditor cap=10
Stage 4: dedup
Stage 5: judge — claude-code (Sonnet 5 high)
                fallback: opencode (GLM-5.2) on primary failure

Usage:

scripts/ultrareview.sh path/to/file.cs
scripts/ultrareview.sh --xml path/to/file.cs
scripts/ultrareview.sh --dry-run path/to/file.cs       # plan check
scripts/ultrareview.sh --no-fallback path/to/file.cs   # disable judge fallback

The Opus-judge fallback is intentional — Claude Code's claude -p backend timed out at 1200s on 200+ findings during the bench. Setting --no-fallback lets you treat a primary judge failure as fatal (useful in CI).

Output filtering

Both modes filter findings via the LLM judge before printing. Verdicts are:

• VALID — kept as-is.
• DOWNGRADE — kept, severity adjusted to new_severity from the judge.
• DUPLICATE — dropped (judge marks the canonical finding it duplicates).
• FALSE_POSITIVE — dropped (hallucination, vague advice, fix doesn't fit the defect, etc.).

The default markdown output groups kept findings by severity. The --xml form preserves the full <code-review-report> schema and adds a <judge-summary> element with verdict counts.

Required config.json entries

These modes hardcode their agent IDs. The default config.json already has all of them defined (most are enabled=false, which is fine — multi-stage modes ignore enabled and look up the entry by id directly):

ID	Where used
codex	ultrareview broad
opencode	ultrareview broad
claude-code	both, plus ultrareview judge
claude-sonnet	superreview judge
opencode-go-minimax	both, discovery + specialist
opencode-go-kimi	ultrareview specialist
opencode-go-glm	superreview specialist, ultrareview probe
opencode-go-qwen37-max	both, broad + specialist
opencode-go-qwen37-plus	both, discovery + specialist

If any are missing the script exits 4 with the list of missing IDs.

When NOT to use multi-stage modes

• Quick diff review → use code-review.sh. Multi-stage adds 5-10× cost.
• Code under 50 lines → judge has nothing to do; use code-review.sh.
• CI without a judge LLM → use --xml from code-review.sh and parse findings yourself.
• Files >2000 lines → split first; even with the judge, the union XML becomes hard to score reliably.

When to Use Which

Pick by role, not by vendor. The default config has Codex (analyst) + OpenCode GLM-5.2 (lateral) enabled, plus the current OC-Go roster; flip claude-code or gemini-cli on in config.json when you want an additional voice.

Situation	Script	Role(s) involved
Code review, security audit	per-agent analyst script (codex-query.sh or claude-query.sh)	analyst — precision, edge cases
Architecture decision, design choice	consensus-query.sh	analyst + lateral — depth + breadth
"Are we solving the right problem?"	per-agent lateral script (opencode-query.sh or gemini-query.sh)	lateral — challenges premises
Bug investigation, root cause analysis	per-agent analyst script	analyst — goes deep into implementation
Exploring alternatives, brainstorming	per-agent lateral script	lateral — cross-domain analogies
High-stakes or irreversible decision	consensus-query.sh	all enabled — reduce blind spots
Agent-to-agent integration (downstream parser)	consensus-query.sh --xml	any — stable structured output

Synthesizing Responses

Agents respond with a shared structure. Compare section by section:

• Assessment vs Assessment: Do they frame the problem differently? A framing difference often reveals the most insight.
• Blind Spots: Union of both agents' blind spots is your risk map.
• Alternatives: Check if either agent proposed something neither you nor the other agent considered.
• Recommendations: Agreement = high confidence. Divergence = investigate the reasoning, not just the conclusion.

Response Patterns

When comparing the two responses, classify the pattern and act accordingly:

• Agreement: Both recommend same approach — high confidence, proceed
• Complementary: Different valid points that don't conflict — combine insights into a richer picture
• Contradiction: Conflicting recommendations — present both with reasoning, let user decide
• Unique insight: One agent caught something the other missed — highlight it, this is often the most valuable output

Prompt Patterns

Architecture Decision (unbiased framing)

scripts/consensus-query.sh "We need real-time updates for ~100 concurrent users.
Updates are server-initiated only. Current stack: [describe your stack].
Latency target: under 500ms from event to UI update.
What approach would you recommend and why?"

Code Review (pipe raw code, let agents form opinions)

cat src/services/auth.py | scripts/codex-query.sh \
  "Review this authentication service. Focus on whatever concerns you most."

Problem Investigation (provide facts, not hypotheses)

scripts/codex-query.sh "Database query returns empty result.
Direct query with same filter returns 5 documents.
[paste query here]
What's happening?"

Environment Variables

• CONSILIUM_CONFIG: Path to a custom JSON config (default: <skill>/config.json)
• CODEX_MODEL: Override Codex model at runtime (default: value from config)
• GEMINI_MODEL: Override Gemini CLI model at runtime (default: value from config)
• OPENCODE_MODEL: Override OpenCode model at runtime (default: value from config)
• OPENCODE_AGENT: Override OpenCode built-in agent (default: plan, read-only)
• OPENCODE_EFFORT: Override OpenCode reasoning effort (default: config effort field, or high)
• CLAUDE_MODEL: Override Claude Code model at runtime (full id or supported alias)
• CLAUDE_PERMISSION_MODE: Override Claude Code permission mode (default: plan)
• CLAUDE_EFFORT: Override Claude Code reasoning effort (default: config effort field, or high if both unset; applies to Opus and Sonnet). Levels: low, medium, high, xhigh, max.
• CODEX_EFFORT: Override Codex reasoning effort (default: config effort field, or high if both unset). Levels: minimal, low, medium, high, xhigh.
• GEMINI_API_KEY: Required for the gemini-cli backend (v1beta model access)
• GOOGLE_GENERATIVE_AI_API_KEY: Required if the opencode backend uses google/... models
• OPENAI_API_KEY: Required if the opencode backend uses openai/... models and OpenCode is not already logged in via opencode auth login
• AGENT_TIMEOUT: Timeout seconds (default: 3600 / one hour)

Prerequisites

• Codex CLI installed and authenticated (codex --version) — for the codex-cli backend
• OpenCode CLI installed (opencode --version) — for the opencode backend. For Zen models (opencode/...) run opencode providers login opencode once; for Google direct models (google/...) set GOOGLE_GENERATIVE_AI_API_KEY; for OpenAI direct models (openai/...) either run opencode auth login and pick OpenAI, or set OPENAI_API_KEY.
• Gemini CLI installed (gemini --version) — for the gemini-cli backend (optional; falls back to direct API)
• Claude Code CLI installed and logged in (claude --version, claude /login) — for the claude-code backend
• GEMINI_API_KEY environment variable — required only when gemini-cli backend is enabled (get key at https://ai.google.dev/gemini-api/docs/api-key)
• Python 3 (for config parsing and Gemini API fallback)