v2.3.0

Added

• audit --path <pathspec>... narrows the survey: each gate reviews only the tracked files both its own --path and the audit's reach, a gate sharing none is skipped, and a literal matching no tracked file is refused before any review starts.

Fixed

• An audit's skip line said "no staged file matches"; an audit reads tracked files, and says so.

v2.2.0

Added

• --standard human-communication flags a pronoun-and-relative-clause standing in for a noun ("what it names", "whatever you ask for") and a mechanism narrated through personified verbs or flourishes ("reaches", "quietly", "for free", "and nothing else").
• --standard agent-communication demands the path, flag, count or error text over "the relevant files", and bans self-narration ("I'll", "note that", "successfully").

v2.1.0

Added

• --standard ai-design-violations ships a numbered list of the visual defaults that mark a UI, slide, diagram, chart or layout as unexamined AI output, for a reviewer to cite by id.

v2.0.6

Added

• --trailer is refused by name, not as an unknown flag: this tool does not carry AI agent commit trailers, and there is no way to add one.

Changed

• Stripping an agent's Co-Authored-By line is announced, naming the removed line and warning that restoring it with --amend can break the organization's audit tooling — no longer a silent rewrite.

v2.0.5

Changed

• attest refuses only a file staged at one version and edited since, not every unstaged edit. Reviewing a subset of the worktree no longer needs a stash.
• The staged scope is stated where a caller reads it: the hook rejection, the spawn line, and the reviewer's brief.

v2.0.4

Fixed

• The severity rules the tool states. A MODERATE is a required fix, and only a MAJOR demands a re-review.

v2.0.3

Added

• npm channel: npm install -g git-agent-verdict installs the same prebuilt binary.

v2.0.2

Changed

• README rewritten for the reader deciding whether to adopt it. No code change.

v2.0.1

Fixed

• await waits with a blocking read. BSD poll does not report a hangup on a FIFO, so it never returned on macos.

v2.0.0

Changed

• A review runs in a process of its own and outlives the caller that started it.
• attest starts a review and returns; it no longer commits.
• One run reviews every gate in declaration order, stopping at the first MAJOR.
• The lock stops at that process. Nothing a reviewer leaves behind can hold the repository.
• One directory per commit holds every log a review wrote, numbered as the gates ran.
• A verb that is told to act names the next command in full.

Added

• await blocks until the review answers and exits on its verdict.
• abort ends a review; the verdicts earlier gates recorded are kept.
• commit creates the commit once every gate has passed.

Removed

• --confirm-running-in-background-shell-with-long-timeout, and both foreground refusals.

Changed

• A review runs in a process of its own and outlives the caller that started it.
• attest and audit wait for that round; killing the caller no longer ends it.
• The claim stops at that process. Nothing a reviewer leaves behind can hold the repo.
• One directory per round holds every log it wrote, named before the first reviewer runs.

Added

• await re-attaches to the round a repo has running and exits on what it decided.
• abort ends one, keeping every verdict already earned.

Removed

• --confirm-running-in-background-shell-with-long-timeout, and both foreground refusals.