本文尚未翻译

本文暂时显示英文原文，中文翻译正在进行中。翻译完成后将自动更新。

原文链接：English Version

Agent Loop Internals

The core orchestration engine is run_agent.py's AIAgent class — roughly 9,200 lines that handle everything from prompt assembly to tool dispatch to provider failover.

Core Responsibilities

AIAgent is responsible for:

• Assembling the effective system prompt and tool schemas via prompt_builder.py
• Selecting the correct provider/API mode (chat_completions, codex_responses, anthropic_messages)
• Making interruptible model calls with cancellation support
• Executing tool calls (sequentially or concurrently via thread pool)
• Maintaining conversation history in OpenAI message format
• Handling compression, retries, and fallback model switching
• Tracking iteration budgets across parent and child agents
• Flushing persistent memory before context is lost

Two Entry Points

# Simple interface — returns final response string
response = agent.chat("Fix the bug in main.py")

# Full interface — returns dict with messages, metadata, usage stats
result = agent.run_conversation(
    user_message="Fix the bug in main.py",
    system_message=None,           # auto-built if omitted
    conversation_history=None,      # auto-loaded from session if omitted
    task_id="task_abc123"
)

chat() is a thin wrapper around run_conversation() that extracts the final_response field from the result dict.

API Modes

Hermes supports three API execution modes, resolved from provider selection, explicit args, and base URL heuristics:

API mode	Used for	Client type
chat_completions	OpenAI-compatible endpoints (OpenRouter, custom, most providers)	openai.OpenAI
codex_responses	OpenAI Codex / Responses API	openai.OpenAI with Responses format
anthropic_messages	Native Anthropic Messages API	anthropic.Anthropic via adapter

The mode determines how messages are formatted, how tool calls are structured, how responses are parsed, and how caching/streaming works. All three converge on the same internal message format (OpenAI-style role/content/tool_calls dicts) before and after API calls.

Mode resolution order:

1. Explicit api_mode constructor arg (highest priority)
2. Provider-specific detection (e.g., anthropic provider → anthropic_messages)
3. Base URL heuristics (e.g., api.anthropic.com → anthropic_messages)
4. Default: chat_completions

Turn Lifecycle

Each iteration of the agent loop follows this sequence:

run_conversation()
  1. Generate task_id if not provided
  2. Append user message to conversation history
  3. Build or reuse cached system prompt (prompt_builder.py)
  4. Check if preflight compression is needed (>50% context)
  5. Build API messages from conversation history
     - chat_completions: OpenAI format as-is
     - codex_responses: convert to Responses API input items
     - anthropic_messages: convert via anthropic_adapter.py
  6. Inject ephemeral prompt layers (budget warnings, context pressure)
  7. Apply prompt caching markers if on Anthropic
  8. Make interruptible API call (_api_call_with_interrupt)
  9. Parse response:
     - If tool_calls: execute them, append results, loop back to step 5
     - If text response: persist session, flush memory if needed, return

Message Format

All messages use OpenAI-compatible format internally:

{"role": "system", "content": "..."}
{"role": "user", "content": "..."}
{"role": "assistant", "content": "...", "tool_calls": [...]}
{"role": "tool", "tool_call_id": "...", "content": "..."}

Reasoning content (from models that support extended thinking) is stored in assistant_msg["reasoning"] and optionally displayed via the reasoning_callback.

Message Alternation Rules

The agent loop enforces strict message role alternation:

• After the system message: User → Assistant → User → Assistant → ...
• During tool calling: Assistant (with tool_calls) → Tool → Tool → ... → Assistant
• Never two assistant messages in a row
• Never two user messages in a row
• Only tool role can have consecutive entries (parallel tool results)

Providers validate these sequences and will reject malformed histories.

Interruptible API Calls

API requests are wrapped in _api_call_with_interrupt() which runs the actual HTTP call in a background thread while monitoring an interrupt event:

┌──────────────────────┐     ┌──────────────┐
│  Main thread         │     │  API thread   │
│  wait on:            │────▶│  HTTP POST    │
│  - response ready    │     │  to provider  │
│  - interrupt event   │     └──────────────┘
│  - timeout           │
└──────────────────────┘

When interrupted (user sends new message, /stop command, or signal):

• The API thread is abandoned (response discarded)
• The agent can process the new input or shut down cleanly
• No partial response is injected into conversation history

Tool Execution

Sequential vs Concurrent

When the model returns tool calls:

• Single tool call → executed directly in the main thread
• Multiple tool calls → executed concurrently via ThreadPoolExecutor
  • Exception: tools marked as interactive (e.g., clarify) force sequential execution
  • Results are reinserted in the original tool call order regardless of completion order

Execution Flow

for each tool_call in response.tool_calls:
    1. Resolve handler from tools/registry.py
    2. Fire pre_tool_call plugin hook
    3. Check if dangerous command (tools/approval.py)
       - If dangerous: invoke approval_callback, wait for user
    4. Execute handler with args + task_id
    5. Fire post_tool_call plugin hook
    6. Append {"role": "tool", "content": result} to history

Agent-Level Tools

Some tools are intercepted by run_agent.py before reaching handle_function_call():

Tool	Why intercepted
todo	Reads/writes agent-local task state
memory	Writes to persistent memory files with character limits
session_search	Queries session history via the agent's session DB
delegate_task	Spawns subagent(s) with isolated context

These tools modify agent state directly and return synthetic tool results without going through the registry.

Callback Surfaces

AIAgent supports platform-specific callbacks that enable real-time progress in the CLI, gateway, and ACP integrations:

Callback	When fired	Used by
tool_progress_callback	Before/after each tool execution	CLI spinner, gateway progress messages
thinking_callback	When model starts/stops thinking	CLI "thinking..." indicator
reasoning_callback	When model returns reasoning content	CLI reasoning display, gateway reasoning blocks
clarify_callback	When clarify tool is called	CLI input prompt, gateway interactive message
step_callback	After each complete agent turn	Gateway step tracking, ACP progress
stream_delta_callback	Each streaming token (when enabled)	CLI streaming display
tool_gen_callback	When tool call is parsed from stream	CLI tool preview in spinner
status_callback	State changes (thinking, executing, etc.)	ACP status updates

Budget and Fallback Behavior

Iteration Budget

The agent tracks iterations via IterationBudget:

• Default: 90 iterations (configurable via agent.max_turns)
• Shared across parent and child agents — a subagent consumes from the parent's budget
• Two-tier budget pressure via _get_budget_warning():
  • At 70%+ usage (caution tier): appends [BUDGET: Iteration X/Y. N iterations left. Start consolidating your work.] to the last tool result
  • At 90%+ usage (warning tier): appends [BUDGET WARNING: Iteration X/Y. Only N iteration(s) left. Provide your final response NOW.]
• At 100%, the agent stops and returns a summary of work done

Fallback Model

When the primary model fails (429 rate limit, 5xx server error, 401/403 auth error):

1. Check fallback_providers list in config
2. Try each fallback in order
3. On success, continue the conversation with the new provider
4. On 401/403, attempt credential refresh before failing over

The fallback system also covers auxiliary tasks independently — vision, compression, web extraction, and session search each have their own fallback chain configurable via the auxiliary.* config section.

Compression and Persistence

When Compression Triggers

• Preflight (before API call): If conversation exceeds 50% of model's context window
• Gateway auto-compression: If conversation exceeds 85% (more aggressive, runs between turns)

What Happens During Compression

1. Memory is flushed to disk first (preventing data loss)
2. Middle conversation turns are summarized into a compact summary
3. The last N messages are preserved intact (compression.protect_last_n, default: 20)
4. Tool call/result message pairs are kept together (never split)
5. A new session lineage ID is generated (compression creates a "child" session)

Session Persistence

After each turn:

• Messages are saved to the session store (SQLite via hermes_state.py)
• Memory changes are flushed to MEMORY.md / USER.md
• The session can be resumed later via /resume or hermes chat --resume

Key Source Files

File	Purpose
run_agent.py	AIAgent class — the complete agent loop (~9,200 lines)
agent/prompt_builder.py	System prompt assembly from memory, skills, context files, personality
agent/context_compressor.py	Conversation compression algorithm
agent/prompt_caching.py	Anthropic prompt caching markers and cache metrics
agent/auxiliary_client.py	Auxiliary LLM client for side tasks (vision, summarization)
model_tools.py	Tool schema collection, handle_function_call() dispatch

Related Docs

• Provider Runtime Resolution
• Prompt Assembly
• Context Compression & Prompt Caching
• Tools Runtime
• Architecture Overview