Read text files

Get sandbox from agent_state, or raise SandboxError.

Resolve relative paths against sandbox work_dir (string resolution only).

Get LLM config for the specified role. role='agent': read from config.llm role='evolve': fields in evolve_agent.llm_config take priority, fall

Find the latest job directory.

Identify spans that correspond to LLM calls.

Install adb from bundled _source/ if not already available. Returns True on success.

Return a standardized tool error payload.

Estimate tokens for diagnostics; never raise from middleware paths.

Send notification via Feishu custom bot Webhook. Config requires: notify: feishu_webhook: "https://open.feishu.cn/open-apis/bot/v2/

Extract a short plain-text summary for overview lines (single line). adb replies are often Markdown (## 回答, numbered sections). The old `. `

Initialize the MCP session, with fallback to HTTP+SSE transport.

Resolve a validated patch path inside the sandbox.

Build assistant turns and attach following tool calls.

Deep merge two dicts, overlay overrides base.

Locate AHE root and prepend its evolve_agent/ dir to sys.path so import tools.file_tools etc. resolve. Returns the added path. Resolution ord

Extract plain text from a message payload.

Find the corresponding trial directory by task name.

Generate a unified diff for UI display.

Generate summary using LLM. Raises: Exception: If LLM call fails.

Read a full text file from the sandbox without sandbox text truncation.

Append newly-seen messages to full trace in agent context.

Restore original trailing newline state.

Find a sequence of lines with progressively more permissive matching.

Send JSON-RPC payload over HTTP+SSE transport.

Write a text file through the sandbox API.

Deep merge a patch dict into the specified agent yaml file.

Compute detailed evaluation statistics from harbor raw job_dir. When k>1, groups trials by task name: a task passes only if ALL k rollouts pass.

Estimate the size of an object in bytes. Args: obj: Object to estimate size for Returns: Estimated

Normalize generation input payload into a dict.

将 explore-agent 产出的 skill 注册到实验副本的 evolve_agent.yaml， 并在实验副本的 evolve_prompt.md 追加 skill 引导。

Write LLM config to environment variables for ${env.LLM_*} references.

E2B aliases reject ``.``; replace with ``-`` to derive a stable alias.

Construct the MCP initialize request payload.

Compact messages using the configured strategy.

Count tokens for the current prompt using configured token counter.

将 agent 的 history 和 tracer 保存到磁盘。

Write current InMemoryTracer snapshot to disk in cleaned format (atomic via tmp+rename).

Extract a short summary of what changes were made from evolve agent output.

Extract Python exception class name from exception.txt content. Strategy: 1. Prefer regex matching standard exception names (ending with Erro

Extract per-turn timing summary from a trace JSON for timeout analysis.

Locate the adb executable, checking common pip script directories.

Get the next request ID.

Group messages into conversation iterations. An iteration is bounded by ASSISTANT messages: [USER or FRAMEWORK](optional) -> [ASSISTA

Process individual SSE events.

Fallback by keeping the newest messages within token limit.

Initialize the MCP connection with proper handshake.

Make a direct HTTP request to the MCP server.

Append the standard install chain (apt + uv + venv + packages) to ``tpl``. The resulting image has ``uv`` available, a fresh Python venv at `

Normalize common unicode punctuation to ASCII equivalents.

Parse nexau.Agent.run() string output into the inner JSON payload. Raises BudgetExceeded when the agent hit max_iterations without calling co

Parse responses that may include SSE-formatted messages.

读取 base agent yaml，应用 patch，写入同目录临时文件并返回。 使用 tempfile 生成唯一文件名，避免并发实验写同一文件的竞争。 临时文件与 base 同目录，确保 yaml 内的相对路径仍能正确解析。 如果没有 patch，直接返回 base_c

Run `adb ask` for one task and return result dict.

Send the initialized notification as per MCP protocol.