Agent Loop & Error Recovery

The core agentic loop: receive message, call LLM, execute tools, iterate, respond.

Agent Loop Flow

1. Persist user message — stored with importance score
2. Auto-route model — classify query complexity (if not overridden)
3. Build system prompt — base prompt + matched skills + known facts
4. Retrieve context — tri-hybrid memory retrieval
5. Iterate (up to max_iterations):
   • Collect pinned old memories + recent messages (deduplicated)
   • Build OpenAI-format message list
   • Call LLM with error-classified recovery
   • If tool calls → execute each, persist results, continue loop
   • If no tool calls OR final iteration → return text response
6. Max iterations reached → return timeout message

Dynamic Iteration Budget

The agent has a built-in request_more_iterations tool that extends the loop budget when the current limit is insufficient:

• Extends budget by 10 iterations per call
• Hard cap prevents unlimited extension (typically 25 total)
• Requires a reason parameter explaining what remains to be done
• Used when the agent has a clear plan but would otherwise run out of iterations mid-task

Error Recovery Strategy

The call_llm_with_recovery method classifies errors and responds accordingly:

Message Ordering Fixup

To satisfy constraints across Gemini, Anthropic, and OpenAI providers, aidaemon runs a three-pass fixup on the message history before each LLM call:

1. Pass 1: Merge consecutive same-role messages (combines tool_calls arrays)
2. Pass 2: Drop orphaned tool results (no matching assistant tool_call) and strip orphaned tool_calls (no matching tool result)
3. Pass 3: Merge again after orphan removal may create new consecutive same-role messages

Tool Execution

During the loop, each tool call receives:

• _session_id — injected automatically for session tracking
• _untrusted_source — flag set for trigger-originated sessions

Stall & Repetition Detection

The agent loop includes safeguards against getting stuck:

• Stall detection — if the same tool is called 3+ times consecutively with similar arguments, the loop breaks
• Repetition detection — detects repeated response text and forces a break
• Hard iteration limit — default 10, extendable to 25 via request_more_iterations

Session Types