ADR 0015 Own agent loop

ADR-0015 — Own agent loop instead of createAgent

• Status: Accepted (supersedes the agent-loop parts of ADR-0003 and ADR-0006)
• Issue: #150 (spike #114, parent #119)

Context

Agents ran LangChain's createAgent with middlewares for usage, budget, approval and the tool call limit. The whole loop was one opaque step: an approval ended the loop and, on resume, the agent ran again from the start — the model was asked again and the decided call had to be recognised by its arguments. A crash could repeat finished tool calls; parallel calls, a defined order of checks at the move boundary and per-agent limits could not be expressed. The spike (#114) showed an own loop on LangGraph handles all of this with stock primitives.

Decision

• Each agent is a compiled subgraph added as one workflow node: input → model → boundary → approval → tool × N (Send) → collect → model … → agent-replyWith.
• Every finished tool call is stored under its callId; nothing finished is repeated. A crash is not a pause: a call running when the process died runs again with the same callId, which tools use as their idempotency key (at-least-once).
• The move boundary order is fixed: tool judges → agent judges (beforeToolCall) → approval → run → afterToolCall (tool, agent); judge points are empty until #123.
• One call per pause; approval goes through an internal interface backed by the pause seam.
• Limits per call of the agent: modelCalls 12, toolCalls 20; hitting one fails the run with agents.<name>.limits.*.
• The loop owns its hooks (usage per call, run budget before a call, tool timeouts); createAgent, the middlewares and the langchain dependency are removed.
• Reasoning attempts (#79) are removed with it; #123 brings the idea back as BeforeAgentAnswerJudge + maxSendBacks. Tern migration 5 drops the attempts column.

Consequences

• Resume and restart continue the same step; the model is not re-asked for a finished turn.
• Tests need scripted models picked by position (#135), never by call order.
• Side-effecting tools must honour ctx.callId; the docs say so.
• A durable checkpointer is required for crash recovery across processes (#121).