v0.4.0

decide runs on OpenAI's Decisions API, its built-in templates work from Go,
and its guides live on a documentation site.

Added

• OpenAI's Decisions API. --provider openai runs templates on
  gpt-6-luna with OPENAI_API_KEY; Go code uses backend.OpenAI. The API
  is in beta, and image templates don't run on it yet.
• Built-in templates in Go. The templates package returns a built-in
  template's questions, such as templates.CommandRisk.Noul("severe"), and
  flags answers as decide run does.
• A documentation site. decide.deepnoodle.ai
  has a quickstart, three tutorials with recordings of real runs, the
  reference, and recipes. The guides in docs/ moved there.
• A Claude Code guide. The
  guide shows how to
  set up the plugin and use /decide:audit and /decide:hunt.

Changed

• command-risk flags deploys and releases. A publish question,
  flagged at 80%, replaces external, which scored routine pushes high and
  was never flagged. --json output has publish in place of external.
• A shorter run summary. decide run ends with one line that tells how
  to see the run again, and lists flagged items only when it judged more
  than one.

v0.3.0

decide audits a codebase for security mistakes, and pays once for each
answer. The plugin drops its reply check.

Added

• decide runs view --top N. Show the flagged items first, then those
  nearest a flag, to read a long run from the top.
• /decide:audit. A plugin skill: decide ranks every function with
  security, and Claude confirms the top 15 by reading them.
• security. A built-in template that asks, in one request per
  function, about SQL and command injection, SSRF, XSS, weak hashes,
  ciphers and random values, and TLS verification turned off.
• An answer cache. decide run and runs resume reuse answers to the
  same question about the same text from the same model name, and ask only
  about what changed, or when a live answer shows the model changed.
• /decide:hunt. A plugin skill: give it a fix commit or pull request,
  and Claude writes one question about the mistake, tests it on the fix,
  sweeps every function, and confirms the top candidates.

Changed

• The plugin checks commands only in bypass mode, and only for severe
  harm. command-risk gains a severe question, flagged at 80%. Local
  cleanup, such as git checkout -- ., no longer asks.

Removed

• reply-check and the plugin's reply check. It flagged about one
  reply in five, mostly honest ones it could not see the evidence for.

v0.2.1

The Claude Code plugin shows what it checked. The CLI is unchanged.

Changed

• The plugin shows its work. The footer counts checks, each checked
  command or result ends with its verdict, and /decide prints a table.
  The status line shows only when a check is off or skipped.

v0.2.0

Decide comes to Claude Code: a plugin that gets Claude a second opinion
before risky commands, on content from outside, and on its own replies.

Added

• command-risk and reply-check. Judge a shell command before it
  runs, and check a coding agent's reply against the tools it ran.
• A Claude Code plugin. /plugin marketplace add deepnoodle-ai/decide
  checks Claude's shell commands, content from outside, and replies, and
  gives Claude a judge tool. Needs Claude Code 2.1.287 or later.

Changed

• prompt-injection catches steering. It counts text that tells an AI
  what to recommend, and hidden no longer counts text that only
  describes hidden text.
• Results fit the terminal. In a terminal, a long score description or
  preview ends with … instead of wrapping. Piped output is unchanged.

v0.1.0

The first release. Decide asks typed questions about your data and gives
answers with probabilities, from the command line or from Go.

Added

• The decide CLI. Run a template on files, folders, JSONL, JSON, CSV,
  text, Markdown, source code, and images. Runs are saved and can resume.
• Nine templates, from sentiment and triage to code-risk and
  prompt-injection. Write your own with decide templates new.
• Diffs. Pipe git diff to judge each changed hunk or function.
• CI. --format prints CSV, a Markdown report, or GitHub Actions
  annotations. --fail-on exits 2 when an item is flagged or matched.
• The Go library. Eval and Pick ask typed questions, and
  decidetest fakes the server in tests. Packages under patterns/, such
  as gate and rank, build decisions from answers.
• Providers. Jev on the TypeSafe API, Clef on Cloudflare Workers AI, or
  any Jev-compatible service.
• Install with brew install deepnoodle-ai/tap/decide, or download a
  binary with checksums.