forge-panel

name	forge:panel
description	Multi-model panel review. Multiple models review independently, then findings are synthesized.
disable-model-invocation	true
argument-hint	[target: path or instruction] [--output path] [--code] [--models m1,m2] [--roles r1,r2] [--review-type type] [--severity level]
context	fork
allowed-tools	Bash, Read

Panel Review

Run a panel review: fans out the same review task to multiple models in parallel, then synthesizes findings.

Usage

/forge:panel [target] [--code] [--models model1,model2]

Arguments

Argument	Required	Description
target	Optional	File, directory, or instruction on what to review (defaults to cwd)
--code	Optional	Switch: use code review framework (default: document review)
--models	Optional	Comma-separated model list (default: all available)
--roles	Optional	Comma-separated reviewer roles (security, performance, architecture, ...)
--review-type	Optional	Review focus: full, security, performance, quick (security/perf need --code)
--severity	Optional	Minimum severity to report: high or critical
--output	Optional	Write result to file instead of conversation (e.g., review.md)

Available models: !forge workflow list-models

Only use models with status ready in the table above. If the default set includes unavailable models, pass --models <ready models> explicitly. If the user explicitly requested an unavailable model, stop and tell them what proxy or credential is missing rather than silently substituting. If no models are ready, tell the user what's missing and stop.

Models Used

Model	Strength	Via
gpt-5.5	Logical problems, systematic review	litellm-openai proxy
gemini-2.5-pro	Balanced analysis, large context	litellm-gemini proxy
claude-opus	Deep architecture, complex reasoning	Direct Anthropic

---

Execution

Step 1: Resolve Target and Flags

Parse $ARGUMENTS into a positional target and optional flags. The target is the first non-flag value (file path, directory, or free-form instruction). Strip any leading @ prefix on the target (Claude Code file reference syntax). If no target is found, default to the current working directory.

Recognized flags (extract from $ARGUMENTS if present):

• --code — switch
• --models <value> — comma-separated model list
• --roles <value> — comma-separated role list
• --review-type <value> — one of: full, security, performance, quick
• --severity <value> — one of: high, critical
• --output <path> — write result to file instead of conversation

Never ask the user to clarify. If $ARGUMENTS contains anything, proceed immediately.

Step 2: Run Multi-Model Review

Execute the panel workflow, forwarding all parsed flags:

forge workflow panel <target> [--code] [--models <models>] [--roles <roles>] [--review-type <type>] [--severity <sev>] --json --cwd "$(pwd)"

Omit any flag the user didn't specify.

Parse the JSON output. The structure is:

{
  "prompt": "...",
  "results": {
    "gpt-5.5": {"response": "...", "error": null, "success": true, "duration_seconds": 45.2},
    "gemini-2.5-pro": {"response": "...", "error": null, "success": true, "duration_seconds": 38.1},
    "claude-opus": {"response": "...", "error": null, "success": true, "duration_seconds": 52.7}
  },
  "successful": 3,
  "failed": 0
}

Step 3: Synthesize Results

Read ${CLAUDE_SKILL_DIR}/resources/synthesis.md for synthesis instructions. If the file is missing, report the actual missing-path problem and stop. Then respond with:

1. Consensus issues (found by 2+ models)
2. Unique findings from each model
3. Conflict resolution
4. Unified priority list
5. Suggested fix order based on dependencies

Output routing: If --output was specified, write the complete synthesis to that path using the Write tool (create parent directories if needed). Print a one-line confirmation: Wrote synthesis to {path}. Do not also print the full result in the conversation. If --output was not specified, print the result in the conversation as usual.

---

Error Handling

• If 1 model fails: Include its error, synthesize from successful models
• If 2+ models fail: Report failure, do not attempt synthesis
• If proxy not available: forge workflow panel skips that model and reports the error in JSON

Requirements

• Forge CLI: forge must be on PATH
• Proxies: GPT-5.5 and Gemini require active proxies (forge proxy create litellm-openai)
• List available models: forge workflow list-models