Sitelet https://github.com/Continuum-AI-Corp/OrcaPromptVault/tree/main/Muse-Code
Skip to content

Latest commit

 

History

History

README.md

Muse Code

captured with OrcaReplay

The file here is an OrcaReplay capture: muse-code-muse-spark-1.3-system-prompt-2026-09-29.md, 24,865 characters, taken 2026-09-29, with its 25 tool schemas beside it.

Meta's terminal coding agent, powered by Muse Spark and installed with curl -fsSL https://dev.meta.ai/install.sh | bash. It identifies itself as "Muse Code powered by Meta Muse Spark" and speaks the OpenAI Responses dialect, with the prompt in instructions rather than in a system message.

Not to be confused with Meta-AI/, which is the consumer assistant. Both run on Muse Spark; they are different products with different prompts, which is why they are different folders.

The tool shape is unlike anything else here

The other 25-tool captures in this archive send twenty-five tools. Muse sends one, of type: "namespace", with the real twenty-five nested inside it — workflow, read_file, search, write_file, edit_file, three *_memory calls, powershell and powershell_input, monitor, three cron_* calls, four goal-and-progress calls, web_search, read_skill, write_todos and the rest. The prompt refers to them as muse.write_file, and in the binary it refers to them as {{tool:write_file}}: the shipped prompt is a template, rendered at launch.

Rendering that template reproduces the captured text exactly, which is how this entry's provenance was checked rather than assumed. The binary it came from is Authenticode-signed by Meta Platforms, Inc., and install.sh installs the same version and channel, 1.4.0-R4302.1.

Two turns, one prompt

Muse sends two model calls per prompt: the agent's turn, and a judge that decides whether a skill should be loaded. Both carry this same 24,864-character prompt; they differ in the context block and in how many of the nested tools they offer — twenty-five for the agent, one for the judge. The file here is the agent's.

How it was reached

Muse fetches a model catalogue with GET /muse-code/models and will not build a turn until that answers, and OrcaReplay's proxy answers 404 to every non-POST because only a POST is ever a model call. So the capture puts a shim in front of the proxy that serves that one GET and forwards everything else untouched — a catalogue is not a model call, and the turn carrying the prompt still goes through the proxy and is recorded there. The shim ships with the capture script, so capture.mjs muse is still one command.

No account was needed. The key is a placeholder, the turn comes back refused, and the prompt travels in the request either way.

Details in docs/CAPTURES.md.