catalyst-probes

Repository removed. As of 14 October 2026, TransmuteLabs/Catalyst no longer exists on GitHub, so the install commands on this page no longer work. We keep the page as a record of what the mod did and will update it if the author publishes it again.

catalyst-probes turns a TOML file into judges for Claude Code. Each [probe.<id>] table in ~/.claude/probes/probes.toml names a trigger, a model and an action; when a main-loop tool call matches, the mod sends the session so far to that model and acts on its one-line verdict, up to cancelling a subagent dispatch. [prompt.<id>] tables in the same file add text to a system prompt section, a tool description or a command description. It is the function-hooks part of Catalyst, the same author’s skills family for disciplined agentic development, and runs without it. Every consultation is a model call and costs tokens. Version 0.1.61 at the time of writing.

What it does

• judge, built in: reviews every Agent and Task call from the main loop before it runs. OK or WARN lets it through; BLOCK, STOP or DENY cancels it with the model’s reason in the tool result.
• idle-watch, built in: with no subagent running and cooldown_min (30 by default) elapsed, asks the model whether to nudge you; a NUDGE verdict becomes a toast.
• form, built in: no model call. It checks briefs, reports, SendMessage texts, git commit and git push commands and, since 0.1.55, file writes through the shell (cp, mv, install, dd, in-place sed and perl, truncate, forced ln -s, redirections and tee) against regular expressions from its table. A refused write is judged after it ran and the file is restored from a backup taken first.
• A failover ladder: when the model provider refuses a consultation, or a subagent’s step, the mod moves to the next model named in the ladder and tells the main loop which model served the step. /catalyst-ladder prints the rungs that are cooling down.
• Your own probes: one [probe.<id>] table plus ~/.claude/probes/<id>/prompt.md. A new id starts as act = "log_only"; nudge and cancel come after you have read its journal.

Install

Needs Claude Code 2.1.287 or later; the author’s supported floor is 2.1.270, stated in the plugin notes, and the type contract the module is written against is 2.1.287.

claude plugin marketplace add TransmuteLabs/Catalyst
claude plugin install catalyst-probes@catalyst

The judge and the idle probe are off until you switch them on in ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_JUDGE": "1"
  }
}

CLAUDE_JUDGE and CLAUDE_IDLE are off when empty; CLAUDE_JUDGE=enforce makes a blocking verdict binding whatever the table says. CLAUDE_FORM, CLAUDE_PROBES (your own ids) and CLAUDE_PROMPTS (the prompt layer) are on when empty and off with 0, off or false. The built-in probes started as ports of “splices”, patches that the ContinuousClaude binary layer applied to the Claude Code executable, and each had a carrier variable (CLAUDE_JUDGE_CARRIER, CLAUDE_FORM_CARRIER, CLAUDE_IDLE_CARRIER) saying which copy ran. The patch copies are gone, so a carrier left unset or empty now means the mod; one set to anything but mod makes that probe refuse every call in its scope, loudly, as a configuration error. Add tables, run /reload-plugins and look for hooks module catalyst-probes loaded in the debug log; the author counts that line as acceptance, not a green claude plugin validate.

A minimal consultant, paired with ~/.claude/probes/brief-check/prompt.md:

[probe.brief-check]
enabled = true
act = "log_only"
rx = "OK|WARN|BLOCK"
models = ["glm-5.3"]
cooldown_min = 60

  [probe.brief-check.when]
  field = "tool"
  in = ["Agent", "Task"]

How it works

The mod hooks tool.call for the main loop only; subagent calls pass through untouched. It reads probes.toml from the global home (~/.claude/probes/, $CLAUDE_CONFIG_DIR/probes or CLAUDE_PROBES_DIR) and from the nearest .claude/probes/probes.toml above the working directory, project keys winning; a project can switch a built-in off with enabled = false. The when predicate (equals, in, matches, count_below, older_than_min and a few more, combinable with all, any and not) tests the tool name, the subagent type, the dispatch prompt, the number of live agents from $.agent.list() and the time of the probe’s last consultation. A custom probe without when never fires.

A consultation is one $.model.complete call per rung of the models ladder until a reply starts with a word from rx (glm-5.3 when no model is set; CLAUDE_JUDGE_MODEL overrides it for the judge). The prompt is <id>/prompt.md (a project’s copy replaces it, prompt.extra.md is appended), then the transcript from $.session.messages() clipped to the last context_chars (24,000 by default, each message cut at 2,000), then for the judge the dispatch itself: tool, subagent type, model and up to 16,000 characters of its prompt. attach_files adds .md and .txt files named in that prompt, up to 90,000 characters. Each consultation therefore costs about the size of your recent transcript plus a reply of up to max_tokens (8,000 by default). A ladder moves on only when the provider refused the rung (no usage and no stop reason, or a throw before any content), never on the content of an answer.

act = "cancel" waits for the verdict before the tool runs; log_only and nudge consult in the background, capped at eight by a counter in the plugin store. Verdicts of cancel probes are memoised by tool, agent and a hash of the prompt for the session, so an identical retry gets the same answer without a new call; the coaching text the mod appends to the communication:L section of the system prompt tells the model to change the dispatch rather than repeat it. If no rung answers, the judge denies (fail_closed is true only for it) and other probes let the call through. /clear and /resume empty the mod’s memory of the session: the verdict cache, the queues and the ladder’s cooldown marks.

The same ladder watches subagents. On agent.spawn the mod records the agent’s declared model and class, and on turn.step it watches each step of the agent for a provider refusal; a refused step is retried on the next rung (failover.default in the table, or failover.main for the main loop), the records say which model served it, and a step served by a model other than the declared one queues one hint for the main loop per agent and rung. Reviewers get their own treatment: a reviewer’s plan drops the rungs that served another reviewer in the last two hours, so the executor and its reviewer do not end up on one model.

Everything lands under the global probes home: <id>/records/mod-<tool_use_id>.json is the full record of a consultation, including the first 2,000 characters of the raw reply (record = false skips it), and <id>/journal.jsonl.shard.* is one index line per record. Prompt rules that apply write prompts/records/applied-<id>.json, and the ladder writes failover/journal.jsonl. The mod makes no network calls of its own; the model call goes through Claude Code.

Prompt rules fire on prompt.section, tool.describe and command.describe. A rule names one target, a mode (append, prepend or replace), its text inline or from <home>/prompts/<text_file>, and optionally when_env with one of the five variables above (any other name refuses the rule). Measured by the author on 2.1.267: 26 sections, 24 tool descriptions and 254 command descriptions in the main loop. A subagent’s system prompt carries one section, so a section rule never reaches it; a tool rule reaches every agent that holds that tool. The mod no longer carries AGENTS.md as a fallback for CLAUDE.md: Claude Code 2.1.277 ships that itself.

Limitations

• Main loop only. Subagent tool calls are not judged; the ladder follows a subagent’s steps, the judge does not.
• The form probe needs its canon, about thirty regular-expression keys in [probe.form] (brief_path, fence, git_msg and so on). The plugin’s notes do not list them; they point to a kit that is not in the repo. Since 0.1.55 a missing required key makes the form judge refuse rather than stay silent.
• A cancel consultation has to finish inside the host’s 10 second hook budget, which the mod cannot raise; set timeout_ms per rung with that in mind.
• The reach figures date from 2.1.267 and have not been re-measured since the floor moved to 2.1.270.