v1.2.0

Features

• Tool calling: the model can now invoke built-in read and bash tools during chat, each gated by a per-call permission prompt.
• Press Esc to interrupt an in-flight reply.

Fixes

• Hardened the read/bash tools and the permission gate against untrusted model output, and tightened probe parsing.
• Tool output is now capped to the on-device context window.
• Model-runtime errors (including context-window-exceeded, which now triggers a compaction retry) surface as clean, readable errors.

v1.0.1

Releases

No user-facing changes in this release.

v1.0.0

Features

• apple-fm now installs on any platform — non-macOS systems install cleanly and degrade gracefully instead of failing, while the on-device model remains available on macOS 26+.
• Helper binaries are now signed and notarized.

v0.1.0

First tagged release.

Added

• Swift apple-fm-helper (--probe / --generate / --session) over an NDJSON
  wire protocol, and a tested, zero-runtime-dependency Node layer.
• CLI (apple-fm): probe, generate (prompt or stdin; --system, --schema,
  --stream, --temp, --max-tokens), and an interactive chat with slash
  commands (/reset, /system, /clear, /compact, /help, /quit).
• Library API: probe, generate, ChatSession, LiveSession, protocol
  helpers, and types.
• Native guided generation. --schema compiles a JSON Schema to a
  GenerationSchema, so the model's output is guaranteed to conform (numeric
  minimum / maximum enforced); structured output can also stream, as JSON
  snapshots.
• Persistent multi-turn chat. chat reuses one on-device
  LanguageModelSession across turns (KV-cache) instead of replaying the
  transcript, and auto-compacts older turns near the context window.

Verified

• Smoke-verified on a macOS 26 / Apple Intelligence machine: probe, generate
  (freeform / guided / streamed), guaranteed structured output, and multi-turn
  chat.

Not yet done

• An automated on-device test in CI (AF-2) and the signed + notarized release
  pipeline (AF-12, verified on the first tagged release) are tracked.