@splitfireai / agent
Code Commits Issues Pulls Cloud Agent CI/CD
development
development main
md 5.17 KB
View source Raw

AGENTS.md

This file provides guidance to coding agents when working with code in this repository.

Commands

CI (.github/workflows/ci.yml, ubuntu + macOS) runs exactly these:

cargo fmt -- --check
cargo clippy --locked --all-targets -- -D warnings
cargo test --locked
  • Single test: cargo test --test acp_integration <name> (or --lib <name> for unit tests in src/).
  • tests/live.rs hits real Onde Cloud and silently skips unless ONDE_API_KEY=app-id:app-secret is set: ONDE_API_KEY=... cargo test --test live -- --nocapture.
  • Enable the tracked pre-commit hook (runs cargo fmt --check on staged .rs files) once per clone: git config core.hooksPath .githooks.
  • Run locally: cargo run -- --setup (store API key), then cargo run (TUI) or cargo run -- --acp (ACP over stdio). SPLITFIRE_YOLO=1/--yolo skips permission prompts; logs go to stderr via RUST_LOG.

Architecture

This crate holds only the music part of the agent. The ACP server, session/turn loop, approval, workspace tools (read_file, write_file, run_command, ...), terminal UI and MCP layer all come from the Onde Agent Platform crates (ed-acp, ed-acp-tui, ed-mcp, from the onde-ed workspace). Look there for anything protocol- or session-related; the README says it is checked out next to this repo.

  • src/profile.rs: SplitFire implements ed_acp::Profile. It is the single wiring point: identity (INFO, env prefix SPLITFIRE), system prompt, toolsets (WorkspaceTools + MusicTools), slash commands, and the lazily started built-in demucs --mcp server (SharedServer; a client-supplied MCP server named demucs wins). Embedding apps use that: the separator passes its own --mcp server, and KaroKowe Jam and KaroKowe Karaoke wrap SplitFire and return no built-in servers.
  • src/prompt.rs: system prompt and slash commands (/analyze, /chords, /practice, /stems only when separation is available).
  • src/tools/: MusicTools is one Toolset exposing six theory_* tools, analyze_audio and list_separations. All are read-only and never prompt for approval. New tools need an entry in definitions(), handles() and describe().
  • src/theory/ (pitch, interval, scale, chord, key, tempo) and src/audio/ (symphonia decode, loudness/tempo/key analysis) are the pure computation behind the tools. The design principle is that the model calls these instead of recalling answers, so results must be exact (correct enharmonic spelling, e.g. C# major has E# and B#).
  • src/lib.rs vs src/main.rs: the crate is both a library (Profile for hosts that embed the agent over an in-memory channel with ed_acp::serve) and the splitfire-agent binary. The cli feature (default) pulls in ed-acp-tui and the MCP test stub; build with --no-default-features to check the embeddable library alone.
  • src/embed/ (behind the off-by-default embed feature) is the SplitFire-only glue a Tauri host needs beside the published ed-acp-tauri crate: embed::fetch_key (SplitFire's /agent/key), embed::splitfire_mcp (the SplitFire API's MCP server), and embed::trim (drop run_command and swap the stems paragraph on platforms that can't start processes). It has no tauri dependency — the app identifier comes in as a &str. The NO_STEMS_START/NO_STEMS_END markers embed::trim::replace_no_stems keys off are defined in src/prompt.rs, which writes the paragraph, so the two can't drift.
  • The binary has three modes: TUI when stdin is a terminal; ACP over stdio with --acp or when stdin is not a terminal; plus --setup / --list-models. The TUI is itself an ACP client that spawns splitfire-agent --acp as a subprocess.

Tests

Integration tests run the real binary against a scripted mock model and a test MCP server (src/bin/mcp_stdio_stub.rs, built from ed-mcp's test-server feature, hence it requires cli). tests/acp_compliance.rs runs the platform conformance suite from ed-acp-testkit plus SplitFire-specific cases; shared helpers are in tests/common/.

Conventions

  • Edition 2024. Per .agents/skills/rust-coding-standard: no .unwrap()/.expect() outside #[cfg(test)], don't silently discard errors with let _ =.
  • registry/ holds the ACP registry submission (agent.json); keep its version in step with Cargo.toml when releasing (both at 1.2.0).

Remembered notes

  • Maven Central search.maven.org/solrsearch is search-index-lagged and often returns 0 results for newly published artifacts (e.g. com.ondeinference:onde-inference 1.3.2 returned numFound=0 for multiple query forms, but the artifact was live). Authoritative sources for a "is X published on Maven Central?" check, in order of truth: (1) https://central.sonatype.com/api/internal/browse/component/versions?filter=namespace:<g>,name:<a> — the modern API, returns version + publishedEpochMillis; (2) https://repo1.maven.org/maven2/<group-path>/<artifact>/maven-metadata.xml — has <latest>/<release>/<lastUpdated>, definitive; (3) curl -I https://repo1.maven.org/maven2/<group-path>/<artifact>/<version>/<artifact>-<version>.pom — confirms the POM is live. Always use User-Agent: Mozilla/5.0 because Xcode-shipped Python 3.9's urllib can JSON-decode-error on some endpoints.