# Changelog
All notable changes to this project are documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [1.2.1] - 2026-06-17
### Fixed
- **Commit hook now uses the configured API key**: the default generation path only
attempted the OpenAI multi-step request when the `OPENAI_API_KEY` environment variable
was set, silently ignoring a key stored via `git ai config set openai-api-key`. As a
result every commit fell through to the local programmatic generator and emitted
filename-derived subjects like `Update <file>`. The client is now built from the stored
configuration (honoring a custom `openai-base-url` too), falling back to the environment
variable only when the config holds no usable key.
- **Local-fallback truncation no longer splits words**: oversized fallback subjects are
trimmed to the last whole word (no more `Update controlle`) instead of a hard mid-word cut.
## [1.2.0] - 2026-06-17
### Added
- **Configurable OpenAI model**: any model string is now accepted (not just the built-in set)
and is verified to exist (via the OpenAI/compatible `models` API) at `git ai config set model`
time — best-effort, so offline users aren't blocked. Unknown models fall back to a safe
tokenizer/context size.
- **Custom OpenAI base URL** (`git ai config set openai-base-url <url>`) — point git-ai at a
local or self-hosted OpenAI-compatible endpoint (e.g. ollama).
### Changed
- **Default model is now `gpt-4.1-mini`** (was `gpt-4.1`) — near-equivalent quality for short,
structured commit messages at lower latency and cost on a per-commit tool.
- **Faster large diffs**: truncation now tokenizes once and decodes a token prefix instead of
re-tokenizing on every binary-search step; parallel diff processing reworked to be both
deterministic and correct.
- **Deterministic commit messages**: identical staged changes now always produce the same patch
(and thus message) — previously affected by `HashMap` iteration order.
- Tightened the commit-message prompts (imperative mood, no trailing period, impact-first,
anti-hallucination), with invariant tests.
- Upgraded all dependencies including async-openai 0.41, git2 0.21, tiktoken-rs 0.12, reqwest 0.13;
pinned the Rust nightly toolchain for reproducible builds.
- Removed dead code (orphaned `ollama`/`client`/`profile` modules and the unused `model::run`).
- Honor the configured `max-commit-length` in the final message selection (was hardcoded to 72).
### Security
- Added `cargo audit` and qlty to CI; hardened all GitHub workflows (least-privilege permissions,
SHA-pinned actions, fixed script-injection) and added datajust-style PR governance workflows.
## [1.1.2] - 2026-06-16
### Fixed
- **Multi-step pipeline broken since v1.1.1**: `collect_diff_data()` dropped git
file/hunk header lines (libgit2 origins `'F'`/`'H'`), so the generated patch had no
`diff --git`/`@@` markers. `parse_diff()` could then no longer split the diff into
per-file sections and collapsed every commit into a single `"unknown"` file, producing
generic commit messages. Headers are now preserved. (#91)
- Annotated an `f32` literal in `model::run()` so the crate builds clean under the latest
nightly (the new `float_literal_f32_fallback` lint).
### Tests
- Added real-git2 regression tests (`tests/patch_test.rs`) asserting headers are preserved
and that `to_patch()` output parses back into the correct per-file sections (never
`"unknown"`).
