third-party-call

name	third-party-call
description	Unified interface for calling external AI tools (Claude CLI, Codex, Gemini) in non-interactive mode. Use for independent code review, idea proposal, design analysis, and other tasks requiring fresh context. Centralizes all third-party tool configuration and permissions.
allowed-tools	Bash, Read, Write

Third-Party Call Skill

This skill provides a unified interface for invoking external AI tools in GAAC workflows. It manages:

1. Tool Detection: Check which tools are available
2. Invocation Wrapper: Consistent interface for all tools
3. Security Controls: Read-only access, sandboxing
4. Output Capture: Structured output handling

Purpose

External tools provide:

• Fresh Context: No conversation history bias
• Independent Perspective: Different model, different thinking
• Specialized Capabilities: Web search (Gemini), deep reasoning (Codex)

Supported Tools

Tool	Primary Use	Detection
codex	Code review, analysis	which codex
claude	Fallback for codex	which claude
gemini	Web research, proposals	which gemini

Scripts

Check Available Tools

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/check-tools.sh"

Run Analysis (Codex/Claude)

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-analysis.sh" \
    --prompt-file ./prompt.md \
    --output-file ./output.md \
    --tool codex \
    --context-files "src/**/*.ts"

Run Web Research (Gemini)

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-web-research.sh" \
    --topic "React state management best practices 2024" \
    --output-file ./research.md

Run Peer Check

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-peer-check.sh" \
    --issue-number 42 \
    --output-file ./peer-review.md

Run Code Review (Independent Scoring)

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-code-review.sh" \
    --issue-number 42 \
    --output-file ./code-review.md

Outputs structured markers: <!-- GAAC_REVIEW_SCORE: NN --> and <!-- GAAC_REVIEW_ASSESSMENT: ... -->

Role-Based Model Configuration

Model configuration is read from gaac.md. The --role parameter determines which model config to use:

Role	gaac.md Key	Default	Purpose
analyzer	gaac.models.analyzer	codex:gpt-5.2-codex:high	Synthesis and analysis
analyzer (fallback)	gaac.models.analyzer_fallback	claude:opus	Fallback for analyzer
checker	gaac.models.checker	claude:opus	Critical checking
proposer	gaac.models.proposer	claude:sonnet	Creative proposals
proposer_secondary	gaac.models.proposer_secondary	gemini:gemini-3-pro-preview	Secondary proposer (Gemini)
code_reviewer	gaac.models.code_reviewer	codex:gpt-5.2-codex:high	Independent code review
code_reviewer (fallback)	gaac.models.code_reviewer_fallback	claude:opus	Fallback for code review

Usage:

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-analysis.sh" \
    --role analyzer \
    --prompt-file ./prompt.md \
    --output-file ./output.md

Model Config Format: tool:model:reasoning (e.g., codex:gpt-5.2-codex:high)

Prompts

Prompt	Purpose
prompts/CODE_REVIEWER_PROMPT.md	100-point scoring rubric for code review
prompts/CHECKER_PROMPT.md	Critical checker for proposer-checker-analyzer debates

Tool Configurations

Codex (Preferred for Analysis)

codex exec \
    -m gpt-5.2-codex \
    -c model_reasoning_effort=high \
    --enable web_search_request \
    -s read-only \
    -o "$OUTPUT_FILE" \
    -C "$PROJECT_ROOT" \
    < prompt.txt

Features:

• Deep reasoning (high effort)
• Read-only sandbox
• Web search capability
• Output to file

Claude (Fallback)

claude -p \
    --model opus \
    --permission-mode bypassPermissions \
    --tools "Read,Grep,Glob,WebSearch,WebFetch" \
    --allowedTools "Read,Grep,Glob,WebSearch,WebFetch" \
    < prompt.txt > output.md

Features:

• Read-only tools only
• Web search capability
• Output to stdout

Gemini (Web Research)

GEMINI_CLI_SYSTEM_SETTINGS_PATH="$TEMP_SETTINGS" \
gemini \
    --output-format json \
    --approval-mode yolo \
    "prompt" > output.json

Settings file:

{
  "tools": {
    "core": ["list_directory", "read_file", "glob", "search_file_content", "web_fetch", "google_web_search"],
    "exclude": ["write_file", "replace", "run_shell_command"]
  }
}

Features:

• Google-powered web search
• Read-only codebase access
• No file modifications

Security Model

All external tools run with:

Capability	Allowed	Enforced By
Read codebase	Yes	Tool config
Search codebase	Yes	Tool config
Web search	Yes	Tool config
Write files	No	Sandbox/tool exclusion
Execute commands	No	Sandbox/tool exclusion

Use Cases

1. Peer Code Review (Phase 6.2)

Uses Codex (or Claude fallback) to review implementation:

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-peer-check.sh" --issue-number 42

Output: PASS/NEEDS_WORK with detailed findings

2. Design Proposal (Research Phase)

Uses Gemini for web-enhanced proposals:

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-web-research.sh" \
    --topic "distributed cache invalidation patterns" \
    --output-file ./proposal.md

3. Independent Analysis (Refinement Phase)

Uses Codex/Claude for synthesizing multiple inputs:

bash "${CLAUDE_PLUGIN_ROOT}/skills/third-party-call/scripts/run-analysis.sh" \
    --prompt-file ./analysis-prompt.md \
    --context-files "./docs/draft/*.md" \
    --output-file ./synthesis.md

Error Handling

Exit Code	Meaning	Action
0	Success	Use output
1	Tool not available	Try fallback
2	Tool execution failed	Try fallback
3	Invalid input	Fix and retry
4	Timeout	Increase timeout or simplify

Timeout Configuration

Tool	Default	Env Variable
Codex	5400s	CODEX_TIMEOUT
Claude	300s	CLAUDE_TIMEOUT
Gemini	300s	GEMINI_TIMEOUT

Integration Points

GAAC Stage	Third-Party Usage
Research	Gemini for web proposals
Refinement	Codex/Claude for analysis
Planning	Codex for design review
Work	Peer-check for code review

Files in This Skill

File	Purpose
SKILL.md	This documentation
scripts/check-tools.sh	Detect available tools
scripts/run-analysis.sh	General analysis wrapper with role-based config
scripts/run-web-research.sh	Gemini web research
scripts/run-peer-check.sh	Quick feedback via external model
scripts/run-code-review.sh	Independent scoring review (mandatory)
prompts/CODE_REVIEWER_PROMPT.md	100-point scoring rubric
prompts/CHECKER_PROMPT.md	Critical checker for debates

Best Practices

1. Prefer Codex - Better reasoning for code analysis
2. Use Gemini for web - Google search integration
3. Always capture output - Don't rely on stdout
4. Set timeouts - Prevent hanging
5. Check exit codes - Handle failures gracefully