Resolve a registry model key into runtime LiteLLM settings.

Load a session. Returns (messages, model) or None.

Process one user turn - may involve multiple LLM calls for tool use.

Save conversation to disk. Returns the session id.

Snip tool outputs first, then summarize old turns if needed.

Handle slash commands. Returns True if command was recognized.

Show a compact header for the tool being executed.

Print tool output, truncated for display.

Truncate long tool results, keeping head + tail.

Synchronize runtime objects with the current config.

Build config from environment variables.

Get OpenAI-format tool schemas for all tools.

List saved sessions with metadata.

Return a warning string if the command looks destructive.

Run a single tool call and add the result to context.

Run multiple tool calls concurrently using threads.

Quick null-byte check.

Quick check: if the first 512 bytes contain null bytes, it's binary.

Handle non-streaming responses.

Handle streaming responses, assembling tool calls from deltas.

Replace old messages with a compressed summary.

Best-effort token tracking for streaming responses.

Generate a compact unified diff.

Track directory changes from cd commands.

Walk dir tree, skipping junk dirs and binary files.

Generate system prompt with current environment context.

Send messages to the LLM and yield response chunks. Yields dicts with keys: - {"type": "text", "content": "..."} for text

Count tokens in a message list.

Clear everything except the system prompt.

Convert to OpenAI function-calling schema.

Strip the provider prefix for display.

Run the tool and return a string result.

Assistant messages with content=None should not crash compression.

Tool outputs longer than threshold should be truncated.

Old messages should be summarized when compress() is called.

Grep should not search inside __pycache__ directories.