call-agent

# Call Agent — Cross-App Peer Communication Send a prompt to another Hermes agent over Fly.io's 6PN private network and get a response. This is **peer-to-peer**: the same skill is installed on every agent in the system, and the agent registry (`references/agents.yaml`) lists who can be called. ## Why This Exists Separately from `delegate_task` `delegate_task` runs in-process — the parent agent waits while a child runs in the same machine, blocking the chat. `call_agent` is **cross-app HTTP**: the called agent runs in its own machine, so the caller stays responsive. Use `call_agent` when the work belongs to a different specialist; use `delegate_task` for fine-grained parallelism within the same agent. ## When to Use - **From the gateway:** delegating code work to the squad (`call_agent squad "..."`) - **From the squad:** relaying a question or completion notice to the user via the gateway (`call_agent gateway "Tell Rishi: ..."`) - Any cross-agent call where the work belongs to a peer with specialized skills or resources ## When NOT to Use - Talking to yourself (loops — the script doesn't currently detect this; use `delegate_task` instead) - Work that can finish in the current agent without specialist help - Trivial conversational responses ## How to Use ```bash # Discover available agents bash /opt/data/skills/custom/call-agent/scripts/call_agent.sh # Call a specific agent bash /opt/data/skills/custom/call-agent/scripts/call_agent.sh squad "Implement X in repo Y" bash /opt/data/skills/custom/call-agent/scripts/call_agent.sh gateway "Tell Rishi: build X complete, ready for review" # Get raw JSON (when you need full response metadata) bash /opt/data/skills/custom/call-agent/scripts/call_agent.sh --json squad "..." # Read prompt from stdin (handy for long prompts) cat plan.md | bash /opt/data/skills/custom/call-agent/scripts/call_agent.sh squad - ``` ## Required Environment The script reads `references/agents.yaml` (relative to the script's parent dir) to find each agent's URL and the env var that holds its bearer token. Each agent's environment must have the bearer tokens for the peers it intends to call. Current registry: | Agent | URL | Bearer env var | |-------|-----|----------------| | `gateway` | `http://hermes-gateway.internal:8642` | `GATEWAY_API_KEY` | | `squad` | `http://hermes-coding-squad.internal:8642` | `SQUAD_API_KEY` | ## Cost Awareness Each remote call loads ~16K prompt tokens (the called agent's full Hermes context). At current model pricing that's roughly $0.005 per call. Bundle related sub-tasks into one call when possible — don't chat back and forth. ## Forward-Compatibility The registry format (`name`, `url`, `description`, `api_key_env`) maps cleanly to A2A Agent Cards. When Hermes ships A2A support (issue #514), the migration is to swap `agents.yaml` for `/.well-known/agent.json` discovery and use the A2A SDK in place of curl. ## Failure Modes - **Network unreachable:** wrong URL, target app stopped, 6PN issue → curl error - **HTTP 401:** bearer token mismatch — check the api_key_env value on both sides - **HTTP 502/504:** target app crashed or busy → check its logs - **Long delay (60s+):** target is doing heavy work — the script's `--max-time` is 10 minutes When in doubt, run `curl http://<agent>.internal:8642/health` first to verify reachability.