Openrouter aider orchestration

--- name: openrouter-aider-orchestration description: | Build multi-model AI orchestrators using OpenRouter with Aider. Use when: (1) "No endpoints found" or "Missing API Key" errors with OpenRouter models, (2) "Argument list too long" when passing large files to Aider, (3) JSON parsing fails due to line-wrapped output from Aider, (4) building cost-effective agent systems with model routing, (5) git merge fails with "untracked working tree files would be overwritten" for .aider. files. Covers model naming conventions, litellm prefix requirements, --read flag for large files, JSON cleaning for programmatic output extraction, and worktree merge handling. author: Claude Code version: 1.1.0 date: 2026-01-22 --- # OpenRouter + Aider Multi-Model Orchestration ## Problem Building AI orchestrators that use multiple models via OpenRouter with Aider involves several non-obvious integration challenges: model naming, API key routing, handling large inputs, and parsing structured output. ## Context / Trigger Conditions - Error: "No endpoints found for anthropic/claude-3-opus" - deprecated model name - Error: "Missing Anthropic API Key" - missing openrouter/ prefix - Error: "Argument list too long" - file content too large for command line - Error: "Invalid control character" in JSON - Aider wraps lines at ~80 chars - Error: "untracked working tree files would be overwritten" for .aider. - merge conflict - Building multi-agent systems with cost-aware model routing ## Solution ### 1. Model Naming Convention OpenRouter requires the openrouter/ prefix when using litellm (which Aider uses internally): ``python # WRONG - causes "Missing API Key" error model = "anthropic/claude-3.5-sonnet" # CORRECT - routes through OpenRouter model = "openrouter/anthropic/claude-3.5-sonnet" ` Current Claude model names (as of 2026): `python CLAUDE_OPUS = "openrouter/anthropic/claude-opus-4" # Was claude-3-opus CLAUDE_SONNET = "openrouter/anthropic/claude-sonnet-4" # Was claude-3.5-sonnet CLAUDE_HAIKU = "openrouter/anthropic/claude-3.5-haiku" # Still valid ` Cost-effective alternatives: `python DEEPSEEK_V3 = "openrouter/deepseek/deepseek-chat" # ~$0.0003/1K tokens GEMINI_PRO = "openrouter/google/gemini-2.0-flash-001" # ~$0.0001/1K tokens QWEN_CODER = "openrouter/qwen/qwen-2.5-coder-32b-instruct" # ~$0.0002/1K tokens ` ### 2. Large File Handling When input files exceed command-line limits (~128KB on Linux), use --read: `python # WRONG - causes "Argument list too long" for large files subprocess.run([ "aider", "--model", model, "--message", f"Analyze this PRD:\n\n{large_prd_content}" ]) # CORRECT - use --read flag to load file subprocess.run([ "aider", "--model", model, "--read", prd_file_path, "--message", "Analyze the PRD file and extract tasks as JSON" ]) ` ### 3. JSON Output Cleaning Aider wraps long lines at ~80 characters, breaking JSON strings: `python def clean_aider_json(output: str) -> dict: """Extract and clean JSON from Aider output.""" # Find JSON block json_match = re.search(r'`(?:json)?\s*([\s\S]*?)`', output) if not json_match: # Try finding raw JSON json_match = re.search(r'\[\s*\{[\s\S]*\}\s*\]', output) if not json_match: raise ValueError("No JSON found in output") json_str = json_match.group(1) if '`' in output else json_match.group(0) # CRITICAL: Fix line-wrapped strings # Aider wraps at ~80 chars, inserting newlines in string values json_str = re.sub(r'\n\s+', ' ', json_str) # Collapse wrapped lines json_str = re.sub(r'"\s*\n\s*"', '""', json_str) # Fix split strings return json.loads(json_str) ` ### 4. Complete Orchestration Pattern `python import subprocess import os def run_agent(issue_id: str, model: str, project_root: str): """Run Aider agent on an issue.""" env = os.environ.copy() env["OPENROUTER_API_KEY"] = os.getenv("OPENROUTER_API_KEY") cmd = [ "aider", "--model", model, # Must include openrouter/ prefix "--openai-api-key", env["OPENROUTER_API_KEY"], "--openai-api-base", "https://openrouter.ai/api/v1", "--message", f"Work on issue {issue_id}", "--no-gitignore", "--auto-commits", "--yes-always", "--no-analytics", ] result = subprocess.run( cmd, capture_output=True, text=True, cwd=project_root, env=env ) return result.stdout ` ## Verification 1. Model works: Agent produces output without API errors 2. Large files work: PRDs >100KB process successfully 3. JSON parsing works: Structured output extracts correctly ## Example ``bash # Cost-effective agent run aider \ --model openrouter/deepseek/deepseek-chat \ --openai-api-key "$OPENROUTERAPIKEY" \ --openai-api-base "https://openrouter.ai/api/v1" \ --read large_spec.md \ --message "Extract tasks from the spec as JSON" \ --