Delegate a generation task to the local Qwen-Coder model via Ollama.

Use this for bulk or routine work where a 14B local model is good enough: generating event templates, headlines, company descriptions, UI polish drafts, boilerplate, or repetitive transformations. The response is returned as raw text — review before applying.

Args: prompt: The task description or content request. system: Optional system message to steer tone / format / constraints. use_brain: If true, prepend .brain/context.md from cwd to the system prompt.

Ask the local model to explain a snippet of code (read-only, no generation).

Useful for getting quick, token-free explanations without consuming Claude's context budget on understanding-only tasks.

Args: code: The code snippet to explain. question: The specific question to answer about the code.

Run several generation prompts in sequence and return all results.

One shared system prompt applies to every item. Prompts are processed serially (Ollama serialises on a single GPU anyway). A failure on one prompt is captured inline as [codebrain error] ... at that index, so the whole batch never aborts.

Returns a single string with per-item delimiters:

--- [0] ---
<result for prompts[0]>

--- [1] ---
<result for prompts[1]>

Args: prompts: List of prompts to run with the same system message. system: Optional shared system message. use_brain: If true, prepend .brain/context.md from cwd to the system prompt.

Apply a targeted transform to existing text — do not regenerate from scratch.

Use this when you have a draft and want it tightened, shortened, rephrased, made more formal, translated, or similar. The system prompt forces the model into transform-mode: it must preserve meaning and structure and only apply the requested change.

Args: text: The existing text to polish. instructions: What transformation to apply (e.g. "shorten to 2 lines", "make tone more formal", "translate to German"). use_brain: If true, prepend .brain/context.md from cwd to the system prompt.

Generate or refresh the <path>.brain summary file for a source file.

Reads the source at path, computes its SHA256, and compares to the existing .brain file's source_hash frontmatter. If they match and force is false, generation is skipped. Otherwise Qwen produces a new brain file (Purpose / Key exports / Collaborators / Gotchas / Conventions), the output is validated against the format spec, and on validation failure one retry with a sharper instruction is attempted before giving up. No partial or broken brain files are ever written.

Format spec: .spec/brain-file-format.md.

Args: path: Path to the source file to summarise. force: If true, regenerate even when the hash matches.

Generate N candidates, let Qwen pick the best, return the winner.

Runs prompt N times (serial — Ollama serialises on single GPU anyway), then does one additional call where Qwen is shown all candidates and asked to return the best one verbatim. Useful for high-variance tasks where a single shot drifts but majority-vote style sampling tightens quality at the cost of N+1 inference calls.

Args: prompt: The task description or content request. system: Optional system message to steer tone / format / constraints. n: Number of candidates to generate (default 3, clamped to [2, 5]). use_brain: If true, prepend .brain/context.md to the system prompt.

Generate with verifier loop — enforces word limits and regex schemas.

Runs codebrain_generate, then checks the output against the requested constraints. On failure, retries with a tightened instruction that names the specific problem. Gives up after max_retries attempts and returns the last output with a [codebrain warning] ... prefix.

Args: prompt: The task description or content request. system: Optional system message to steer tone / format / constraints. min_words: Minimum output word count (None = unbounded). max_words: Maximum output word count (None = unbounded). must_match: Regex pattern the output must match (re.search semantics). max_retries: Max retry attempts on verification failure (default 2). use_brain: If true, prepend .brain/context.md to the system prompt.

Seed .brain/context.md for a repo — one-time setup before scanning.

Detects the stack (python / js / ts / rust / go / java) from marker files, counts source-file extensions, asks Qwen for a short overview, and writes .brain/context.md with a pre-populated template. The user is expected to edit the ## Notes for Claude section afterwards. Idempotent: existing context.md is not overwritten unless force=True.

Args: root: Directory to initialise. force: If true, overwrite an existing .brain/context.md.

Scan every source file under root and generate/refresh its .brain file.

Walks the directory tree, filters by file extension, prunes excluded directories, and runs codebrain_scan_file on each match. Hash-gated: unchanged files skip the model call. Per-file failures do not abort the batch — they are reported at the end.

Defaults:

• extensions: .py .js .ts .tsx .jsx .java .go .rs

• exclude_dirs: .git .venv venv node_modules pycache dist build target

Args: root: Directory to scan recursively. force: If true, regenerate every brain file even when source hash matches. extensions: Override default source extensions (e.g. [".py", ".rb"]). exclude_dirs: Override default directory-name exclusion list.

Report which Ollama models are available locally.

Call this to verify the local backend is reachable and discover which models the user has pulled.