feat: send loop nudge message to agent when loop is detected

[avoidwork]

Summary

When the agent enters a repetitive loop, the system detects it via turn hash tracking but only displays a UI nudge ("You're looping.") to the user. The agent itself never receives feedback that it is stuck, so it continues looping.

Current Behavior

1. src/agent/react.js tracks turn hashes in a sliding window (default size 20)
2. When a duplicate hash is found, it emits a loop_detected event
3. The TUI (src/tui/app.js:349-359) handles this by appending "You're looping." to the last assistant message display
4. The agent continues its loop — it never receives a message telling it to try a different approach

Desired Behavior

When a loop is detected, inject a nudge message into the conversation that the agent can see. This message should:

• Be sent to the agent (not just displayed to the user)
• Inform the agent it is in a repetitive loop
• Encourage the agent to try a different approach

Example nudge: "You are in a repetitive loop. Try a different approach."

Config Design

Two new fields under agent in config.yaml, with corresponding environment variable overrides:

• loopMsg (string, optional) — The nudge message to inject when a loop is detected. If omitted, use a sensible default.
  • Env var: AGENT_LOOP_MSG
• loopLimit (number, optional, default: 5) — Maximum number of loop nudge messages to inject before stopping. After this limit, the system should stop injecting nudges (but may still detect loops).
  • Env var: AGENT_LOOP_LIMIT

Example config:

agent:
  recursionLimit: 1000
  autoContinueLimit: 1000
  nodeTimeout: 600000
  turnHashWindow: 20
  turnBufferMax: 64
  loopMsg: "You are in a repetitive loop. Try a different approach."
  loopLimit: 5

Technical Notes

• Loop detection lives in src/agent/react.js (turn hash tracking, checkTurnHash function around line 265)
• The loop_detected event is emitted via callback({ type: "loop_detected" }) at line 269
• The TUI currently only handles this for display purposes (line 349-359 in src/tui/app.js)
• The nudge should be injected into the conversation messages before the next agent turn, not just displayed in the UI
• Consider whether the nudge should be a user message or a system message
• The turnHashWindow and turnBufferMax are configurable via config.agent.turnHashWindow and config.agent.turnBufferMax
• Env var resolution follows the existing pattern in src/config/loader.js — container keys like agent are dropped, so agent.loopMsg resolves to AGENT_LOOP_MSG

Decisions

• Nudge message is configurable: Yes, via agent.loopMsg / AGENT_LOOP_MSG
• Max nudge count: Yes, via agent.loopLimit / AGENT_LOOP_LIMIT, default 5
• Include repetition context in nudge: No