Ensure the initialize response conforms to expectations.

Add an MCP server configuration.

Archive change_manifest.json from experiment root to the iteration evolve directory.

Build the evolution agent query. Provides concise structured info: evaluation overview, task classification, cross-iteration changes, debugge

Extract patch from explore_agent_patch. If not explicitly specified, inherits api_type/reasoning/tool_call_mode from evolve_agent.

Compact messages to reduce token usage. Args: messages: Full message history Returns: Compacted message list

Extract a cleaned trace dict from normalized observations.

Extract an assistant message from OpenAI/Anthropic style outputs.

Launch harbor evaluation process without waiting. Returns (proc, started_after). Use wait_for_harbor() to wait for completion and get the job_dir

Load config file with _base inheritance chain support.

Normalize an object by converting complex types to serializable forms. Args: obj: Object to normalize max_de

Normalize object to fit within specified size constraints. Args: obj: Object to normalize max_depth: Maximum

Lazy initialization for OpenAIResponsesAggregator.

从 git diff 提取变更信息

Read CSV files

Regenerate iteration_scores.md from iteration_scores.yaml (or provided data dict).

Recursively resolve ${VAR_NAME} in config values with environment variables. Unmatched ones are kept as-is.

Start harbor evaluation and wait for completion. Results are written directly to iteration_dir.

Save the evolve agent's final output to the iteration directory.

Save the evolve agent's full conversation history and cleaned tracer to the log directory.

Save variant evolve outputs to evolve/, and selection+stats to next iteration's input/.

Check if compaction should be triggered. Args: messages: Current message history current_tokens: Current token count

After evolve agent runs, append change summary to evolution_history.md.

Append current iteration scores to iteration_scores.yaml and regenerate .md.

Wait for a previously launched harbor process to finish and return its job_dir.

Add line numbers to content.

Create new ModelCallParams with fallback provider config. RFC-0003: 创建使用备用 provider 的新 params 不修改原始 params，而是创建副本并替换 llm_config 和 cl

Apply replacements in reverse order to preserve indices.

Build one TaskAnalysisJob per task by scanning trial directories.

Construct nexau.Agent from agent_config.yaml with LLM settings patched in.

Build an SDK client from LLMConfig. RFC-0003: 根据 api_type 创建对应的 SDK client

Build the harbor CLI command list.

Match the iteration hint semantics used in executor loop.

构建源码探索 agent 的初始 query。

Format a Codex-compatible success summary.

Replicate executor token limit hint messaging.

Construct the replacement tool output. Args: original_output: The raw tool output. truncated_text: Line-truncated ver

Build a concise verifier output section to include in the debugger query.

构建 Web 调研 agent 的初始 query。

Try exact string replacement. Returns result dict or None if no match.

Try flexible (whitespace-insensitive) replacement. Matches content by stripped/trimmed lines, preserving original indentation.

Try regex-based flexible replacement. Tokenizes the search string and allows flexible whitespace between tokens.

Calculate replacement using multiple strategies. Tries: exact -> flexible -> regex

Validate the patch begin/end markers.

Check Python syntax of all .py files under tools/ and middleware/.

Split messages into chunks, summarize each, then merge. Args: messages: Messages to summarize. chunk_token_limit: Max

Compute the new content that would result from adding a memory entry.

Compute the replacements needed to apply update chunks.

Continuously consume SSE messages from the fallback transport.

Convert GitHub blob URL to raw URL.

Decode sandbox file content while preserving non-UTF-8 text as latin-1.

深度合并两个 dict，overlay 覆盖 base。

Delete a file through the sandbox API, but reject directories.

Derive the new file content after applying update chunks.

根据文件路径推断所属组件

Detect file encoding with fallback to utf-8 (via sandbox read).

Detect visual file type based on extension.

Detect file type based on extension.

Detect the line ending used in content.

Detect line ending used in content.

Ensure proper newline separation before appending content.

Escape special regex characters.

Estimate the JSON string length of an object. Args: obj: Object to estimate length for visited: Set of visit

Execute the MCP tool asynchronously.

Extract the primary text field from a dict-style tool output. Returns a ``(key, text)`` tuple. If *output* is not a dict or does not

Extract key error lines from pytest trace (assertion details starting with E).

Extract error type name from pytest traceback.