Skip to content

Claude Code

The decide plugin gives Claude a second opinion. It adds two skills you run yourself: /decide:audit audits a repository for security bugs, and /decide:hunt finds where a fixed bug was repeated. It also checks, in the background, what Claude reads and runs. For what each check asks, what it sends, and its options, see the plugin’s README.

You need Claude Code 2.1.287 or later, decide 0.5.0 or later, and a key for a decision model.

  1. Install decide and set your key:

    Terminal window
    brew install deepnoodle-ai/tap/decide
    export TYPESAFE_API_KEY=...

    To use Workers AI, OpenAI, or Gemini instead, set that provider’s variables, as Providers shows. Set them where Claude Code will see them, such as in your shell profile. Check that decide works:

    Terminal window
    echo "I love it" | decide run sentiment
  2. In Claude Code, add the marketplace and install the plugin:

    /plugin marketplace add deepnoodle-ai/decide
    /plugin install decide@decide
  3. Restart Claude Code. If it can’t find decide, such as when you start it from an app rather than a terminal, set decidePath in /plugin to the full path from which decide.

To update later, run brew upgrade decide, then update the plugin:

Terminal window
claude plugin update decide@decide

Then restart Claude Code.

You don’t need to do anything for these:

  • Content from outside is checked. After Claude reads a web page, search results, an MCP tool’s result, or gh, curl, wget, or http output, decide checks whether it tries to take over an AI agent or hide text from you. If it does, Claude reads a warning beside the result, and you see a toast.
  • Severe commands are checked when nothing else would check them. When Claude Code would run a shell command without asking anyone, as in bypass mode, decide checks whether it could do severe harm that’s hard to undo, such as deleting a home folder or force-pushing a shared branch. If so, Claude Code asks you, with decide’s reason. In auto mode, its own classifier judges commands, and decide stays out. Set commands to deny or off in /plugin to change this.
  • Claude can ask its own questions. The judge tool, mcp__decide__judge, lets Claude ask typed questions about up to 200 items, such as which of these issues are bugs. It reads text only. For images, PDFs, audio, or video, Claude runs a template on the files instead, such as decide run invoice-check invoices or decide run code-risk diff.png --provider openai.

The footer shows a count, such as decide 12 checked, 1 flagged, and /decide shows the last 15 of the session’s checks in a table. A ⚠ in the status line means some actions went unchecked or a check is off, and /decide says why.

Run /decide:audit from the repository you want to review:

/decide:audit
/decide:audit internal/api cmd/server

Name paths to audit only part of the repository. Claude then:

  1. Picks the server code, leaving out tests, generated and vendored code, and examples, and counts the functions. Over 5,000 functions, it asks you before going on.
  2. Runs decide’s security template on every function. In one request per function, it asks about SQL and command injection, SSRF, XSS, weak ciphers, hashes and random values, and TLS checks turned off.
  3. Takes up to 15 of the highest-scoring functions, those with an answer of 30% or more, and reads each one, with its callers, in parallel. If none scores that high, it says so and stops.
  4. Reports a table of findings, each confirmed or suspected, with the reason, and counts the ones it dismissed.

decide’s answers show where to look. A finding is confirmed only after Claude reads the code. When it’s done, ask Claude to write a failing test for each finding, or to fix them. If more functions scored 30% or more, ask it to read the next 15, which costs no new requests.

The audit doesn’t cover path traversal, authorization, or deserialization, and it can miss a bug that spans several functions.

When you fix a bug, the same mistake is often somewhere else. Give /decide:hunt the fix, as a commit or a pull request:

/decide:hunt 87
/decide:hunt a1b2c3d
/decide:hunt

With no argument, it uses the last commit. Claude then:

  1. Reads the fix and picks the one mistake most likely to be repeated.
  2. Writes one yes-or-no question about that mistake.
  3. Tests the question: it must match the code before the fix and not match the code after it. If it doesn’t, Claude rewrites it.
  4. Asks the question of every function, and reads up to ten of the highest, those that score 30% or more.
  5. Reports which are the same bug and which aren’t.

The question is saved as a template, such as .decide/templates/bug-87, and Claude offers to commit it. Committed, it can check later changes for the same bug, such as in CI:

Terminal window
git diff main | decide run bug-87 --each function --fail-on matched

Each audit and hunt is a saved decide run. To read one again in your terminal:

Terminal window
decide runs # list runs
decide runs view <run-id> --top 15 # flagged first, then nearest a flag

Answers are cached, so running an audit again asks only about functions that changed. The background checks are saved separately:

Terminal window
DECIDE_HOME=~/.decide/agent decide runs

Each check sends the text it checks to the decision model’s provider, with your key:

  • each shell command Claude Code would run without asking anyone
  • each outside result the content check reads
  • with an audit or a hunt, every function it covers

The plugin saves its runs and answers on your disk, under ~/.decide/agent/runs. See What it sends and keeps.

decide is Apache 2.0. Made by Deep Noodle, after some very hard thinking.