№ 0701Skill
pi-muse: Muse Code harness on Pi
A Pi fork that runs Muse Code's system prompt and tool interface on Pi's agent loop and TUI, as an open, hackable harness for Muse Spark 1.3.
# pi-muse A fork of [Pi](https://github.com/earendil-works/pi) that runs Meta's **Muse Code** harness interface on top of Pi's agent loop, TUI, session, and provider layers. ## Why this exists Meta's **Muse Spark 1.3** is widely discussed as a "benchmaxxed" model. Meta co-trained it inside Muse Code and locks its best behavior to that harness, so the model reaches its advertised performance mainly when it is driven by Muse Code's system prompt and tool interface. This project is an alternative, open, hackable harness for the same model: it keeps Pi's runtime and swaps in Muse Code's system prompt and tool contract, so you can run Muse Spark 1.3 (Meta Model API or the OpenCode Go gateway) without depending on Meta's closed CLI. *(The "benchmaxxed"/lock-in framing is the maintainer's motivation, not a statement by Meta.)* ## Status The harness interface is captured from a live Muse Code 1.2.1 session and verified by a traffic-interception harness that runs the real `muse` CLI and `pi-muse` side by side, replays the same canned response to both, and diffs their outbound requests: | | | |---|---| | System prompt | byte-identical (40,445 chars, sent in the `instructions` field) | | Tool surface | **29 / 29** tools, identical order, 0 description diffs, 0 parameter diffs | | Per-session `developer` message | byte-identical, all sections | | Request parameters | all core checks pass (`model`, `max_output_tokens`, `store`, `stream`, `reasoning`, `include`) | | Round trip after a tool call | weighted fidelity 100%, tool-result strings equal | Reproduce it with `npm run muse:parity`. Not implemented: an OS sandbox, deterministic replay, hooks/MCP/plugins. ## Muse Code fidelity The interface was captured from a live Muse Code 1.2.1 session (`muse-spark-1.3-contributor`) rather than reconstructed from docs: the runtime `instructions`, the `muse` namespace tool group, and the per-session `developer` message. The captured prompt is mirrored at https://github.com/souta-lab/muse-code-system-prompt. | Area | Status | |---|---| | Runtime system prompt | Byte-identical to the capture (40,445 chars, `muse.<tool>` refs included) | | Tool surface | 29/29 tools; 0 description diffs; 0 parameter/`strict` leaf diffs; identical order | | `read_file`, `write_file`, `edit_file`, `search`, `bash`, `bash_input`, `read_memory`, `add_memory`, `edit_memory`, `workflow`, `write_to



ChatForm
Tgmlabs