Skip to content

providers.md

Latest commit

 

History

History
36 lines (26 loc) · 4.33 KB

providers.md

File metadata and controls

36 lines (26 loc) · 4.33 KB
Raw

Provider Notes

How git-sc launches the providers that need special handling, and what changes per platform. Which providers run, in what order, and with which model or account is set in configuration.md.

Grok CLI

The Grok provider drives the Grok Build TUI (grok, xAI). That CLI is a full coding agent — plan mode, cross-session memory, web search, and tool execution are all on by default — so git-sc constrains it to behave as a single-turn pure function:

Flag Why
--output-format plain Headless text output instead of the interactive TUI
--sandbox read-only Forbids filesystem writes and network, like Codex's sandbox
--no-plan / --no-memory Blocks plan mode and cross-session memory (both default on)
--disable-web-search Cuts web fetch/search
--max-turns 1 Stops any tool loop after one turn
--verbatim Prevents the CLI from rewriting the prompt
--prompt-file <temp file> Avoids ARG_MAX limits and cmd.exe metacharacter issues for large diffs
  • Model: resolved as model on the step > [models].grok > empty (defer to grok's own default). When non-empty, the ID from grok models (currently only grok-4.5) is passed as -m "<id>". The shipped default for [models].grok is left empty so a cheaper model added later is picked up without a git-sc release.
  • Availability: the Grok CLI ships bundled with cmux at /Applications/cmux.app/Contents/Resources/bin/grok. If grok is not on PATH, this step is skipped and the chain moves on to the next provider.

Apple Intelligence

Apple Intelligence provider uses fm-rs (Rust bindings for Apple's Foundation Models framework) for fully on-device inference. No API key or network connection is required.

  • Requirements: macOS 26 (Tahoe) or later, Apple Silicon, Apple Intelligence enabled in System Settings
  • How it works: With Apple Intelligence enabled (default on macOS), git-sc calls Foundation Models directly via fm-rs. A LanguageModelSession is created with commit-message-specific instructions for each generation. The instructions are built from the resolved prefix type, so prefix_type = "none" / "bracket" / "emoji" and auto-detection from recent commits are respected instead of always forcing Conventional Commits.
  • Context window: the on-device model holds 4096 tokens, which is far smaller than every other provider. git-sc measures the prompt before generating and, when it does not fit, rebuilds it from a compacted diff that keeps the full list of changed files and truncates the bodies. A warning is printed when this happens, because the resulting message was written from a partial view of the change. If even the compacted prompt does not fit, the run moves on to the next provider instead of failing the commit.
  • Timeout: bounded by provider_timeout_seconds (default 60), the same setting the CLI providers use.
  • Failures: a failure caused by the prompt itself (context size, safety guardrail, refusal, unsupported language) does not put the provider into cooldown — only failures that mean the model is currently unusable (assets not downloaded, rate limited, timed out) do. Typed failure classification requires building against the macOS 27 SDK or later; on macOS 26 all failures are treated as provider failures.
  • Build: cargo build --features apple-ai (automatic with make build / make install on macOS). Building on macOS 27 or later additionally enables Foundation Models 27 features (exact token counts, typed errors, per-response token usage); macOS 26 remains supported.
  • Cross-platform: On Linux/Windows, Apple Intelligence is not available and is automatically skipped

Platform Notes

  • Windows: the Antigravity CLI (agy) provider is skipped with an explicit error. All providers launch through cmd /C on Windows (to support npm-installed .cmd shims), but cmd.exe cannot safely receive a multi-line diff prompt as a command-line argument, so passing it would corrupt the command (and is a known command-injection vector class, CVE-2024-24576). The fallback chain simply moves on to the next provider. Providers that read the prompt from stdin or a temp file (codex, claude, opencode, grok) are unaffected.