Decision Tree Claude Code mod: a Claude Code decision tree of options

Decision Tree is a community mod by Ray Amjad (ray-amjad), part of his claude-code-mods collection. When Claude ends a reply by asking you to choose something, the mod builds a Claude Code decision tree in a side pane. The pane shows the question in one plain line, what you need to know to decide, and one box per option with what it means, what it leads to, and short pros and cons. The recommended option is drawn in green with a one-sentence reason.

The last box is "Other": type your own answer and it joins the options. Each box has an Explore button. Explore forks the real session at the message that asked the question, tells the fork you picked that option, and shows what it would do, in up to four steps, and the next decision it would need from you. You can go deeper from there, see the whole tree with /explain map, or open a branch as a live session.

The install block above adds the author's own marketplace and installs the plugin from it. Restart Claude Code or run /reload-plugins, then keep working as normal. The pane appears after a reply that ends in a real question.

After each of the main agent's replies the mod runs a regex on the end of the reply: a question mark in the last paragraph, or two or more numbered or lettered options. Most turns stop there. If it matches, a small model confirms that you have to decide something, and then $.model.fork explains the choice as JSON for the pane. The gate, from hooks/register.jsx:

const r = await $.model.complete({
  model,
  system: 'You read the end of an AI coding assistant reply. Answer YES if it ends by asking ' +
    'the user to decide, choose, or supply something before work can go on. ...',
  prompt: answer.trim().slice(-2000),
  maxTokens: 5,
  timeoutMs: 20000,
})
return r.isAnswered && /^\s*yes/i.test(r.text)

Exploring an option runs claude -p <what-if prompt> --resume <session> --fork-session --resume-session-at <message id> --output-format json through $.process.run, with a 5-minute timeout. To find the message id, the mod reads the session transcript under ~/.claude/projects and looks for the assistant message that ends with the question. If it cannot, it forks from the latest point and logs that. The tree lives in the mod's memory, so a new question, a restart or a reload clears it.

There is one option, gateModel (default haiku), the model that answers YES or NO. Commands: /explain opens the pane for the latest question, /explain b explores option B (A to H), /explain other <answer> explores your own answer, /explain map shows the whole tree, /explain back and /explain top move up, and /explain continue opens the current branch as a live session. For anything else, edit hooks/register.jsx, for example the regex in looksLikeQuestion or the JSON shape in SHAPE.

We read every file the plugin can execute before running it: plugin.json, hooks/hooks.json, hooks/register.jsx (about 430 lines, readable source) and the test file. There is no package.json, install script, bundled code or binary; the two PNGs are README images. We found:

• Model calls: $.model.complete sends the last 2,000 characters of a reply to the gate model, and $.model.fork replays the whole conversation to write the explanation. Both go through Claude Code's own model access, with your tokens and no other host.
• Runs processes ($.process.run): claude -p ... --resume ... --fork-session for each explored option, with --allowedTools=Read,Grep,Glob and --disallowedTools=Bash,Edit,Write,NotebookEdit. When HERDR_ENV=1 it also runs herdr pane split and herdr pane run for Continue branch. All are argv arrays, no shell.
• Files: reads the session transcript .jsonl ($.fs.read, limited to 4 MiB). It writes no files itself, but each branch run leaves its own session file in your project history.
• Environment: reads HOME, USERPROFILE, HERDR_ENV, DECISION_TREE_WAIT and DECISION_TREE_CHILD. The last one is passed to branch runs so they do not explain their own replies. Nothing is sent anywhere.
• Hooks: classic.Stop only records the transcript path and passes the event on. turn.complete never changes the result. Nothing blocks or rewrites a tool call or prompt.
• Clipboard and prompt text: $.ui.copy puts a claude --resume <id> --fork-session command on the clipboard. The fork prompts are written by the mod and aimed at the model only to produce JSON; we found no hidden Unicode or instructions aimed at a reader.
• No network, no telemetry, no daemons, no changes to settings or shell files.

claude plugin validate reported no errors and no warnings from the plugin, only a non-blocking "no icon" note for the plugin directory. The author's 2 tests (the /explain command is registered in interactive sessions and not in headless ones) passed on Claude Code 2.1.307. The tests do not exercise the fork or the pane. It is MIT licensed, with a LICENSE file in the repository.

Checked with Claude Code 2.1.307 on macOS (validate and tests only; we did not record a session). The branch feature relies on --resume-session-at, which the author calls an undocumented Claude Code flag, tested on 2.1.296. If nothing happens:

• The reply must end with a question the gate model judges to need a decision; "anything else?" does not count.
• In a terminal narrower than about 144 columns, run /explain to open the pane.
• In claude -p or SDK sessions the mod does nothing unless DECISION_TREE_WAIT=1 is set.
• A failed branch shows the error in the pane; try the Explore button again.
• Branches take roughly 20 to 60 seconds and cost a full turn each.