Run a self-contained subtask on a LOCAL model and return its text.

Good for: boilerplate, unit tests for a given file, docstrings, log triage, rewrites, first drafts, mechanical transformations. Pass the files it needs in context_files (absolute paths, ~90k chars total). model_tier="fast" is ~1.4x quicker for simple jobs; think=true adds reasoning for tricky ones (slower). Output is saved to a file whose path is returned, so ask for long outputs freely.

NOT for: final correctness decisions, reasoning about the whole repo, tasks needing more than ~32K tokens of context, or anything you must get right without checking.

Second-opinion code review of a diff by a LOCAL model; returns structured findings.

Give either diff text, or git_range and let the server run git diff in cwd (defaults: cwd = current dir, range = HEAD i.e. all uncommitted changes). Examples: git_range="--staged", git_range="main...HEAD". focus narrows the review ("concurrency", "error handling"). Findings carry file, line, severity (critical/major/minor/nit), summary and suggestion.

Use it as a first pass before or alongside your own review, then verify each finding against the code — it is a cheap second opinion, not a verdict.

Summarise large files, logs or text with LOCAL models (map-reduce; megabytes are fine).

Use instead of Read when a file is long (logs, dumps, generated code, long docs) and you need the gist or an answer to a specific question ("what errors occurred after 14:00?"). Accepts absolute paths and/or inline text. Concrete details (identifiers, numbers, errors) are preserved. Full summary is also saved to a file.

NOT for short files (<200 lines): just Read them.

Build or refresh the LOCAL semantic search index for a directory tree.

Incremental: only files whose mtime/hash changed are re-embedded. Stored in the server's data dir (~/.cache/ollama-agent), not in the repo. Run once per repo; search_code refreshes it automatically afterwards. include_globs defaults to common source/doc extensions.

Semantic search over a codebase by MEANING using local embeddings.

Ask in natural language: "where do we retry failed uploads?", "code that parses the config file". Returns the best-matching chunks as path:start-end plus a snippet. Refreshes the index first unless refresh=false.

NOT for exact identifiers or strings — use Grep for those.

Show the active local profile, which model serves each tier, and what Ollama has loaded.

Call this when a local tool is slow or fails, or before a batch of local work, to see VRAM placement and whether an unrelated model is loaded and competing for the GPU.