Compact model context

MCP clients can call get_model_context with a current modelId and revision from list_models. Native and ACP Canvas chat attach the same facts automatically after a matching capture. Source-only and unsupported capture states explain which facts are unavailable. Reading the tool never starts a capture or executes a model.

The packet contains module paths and aliases, complete retained per-module call lists, unique recursive registered parameter element counts, ordered tensor leaves and shapes, per-call modes, provenance, unknowns and capture omissions. Counts include frozen registrations and must not be summed across overlapping modules. These facts describe the recorded path; they do not establish numerical or training correctness, or verify that current inputs, weights and external dependencies match.

The default limit is 4,000 UTF-8 bytes. Use an exact module id, path, qualified type or invocation id as focus; maxBytes accepts 2,000–100,000. Omitted module/call counts stay explicit. Use focused calls or inspect_capture for omitted details. Source changes, an evicted capture or a newer unsuccessful attempt cannot borrow older successful measurements. MCP awaits the host’s existing freshness guard, which may recheck source/setup metadata.

Canvas excludes live facts from draft and historical selections. With available compact facts, its general graph preload has an 8 KiB ceiling instead of 64 KiB; continuation still retrieves omitted records. Human source review and explicit execution controls remain in place. Payload limits do not establish a measured agent speedup or token saving. See source editing.