Loop Detective Claude Code mod: catch Claude Code repeated failures

Loop Detective is a community mod by YeonwooSung (YeonwooSung), part of the my-claude-code-mods collection. It watches for Claude Code repeated failures: the same Bash command, run again and again, failing the same way. After three identical failures it opens a case and tells you in three places: the status line reads loop: 1 active · max 3 failures, a toast names the case, and /loop-report prints the evidence.

A report lists the command, the working directory, the failure line, the call numbers and, when Claude edited files between the retries, the paths it edited. A case with at least two such edit-and-retry steps is labelled edit-retry; otherwise it is retry. The report ends with a hint such as "compare inputs and environment; formulate a different hypothesis before retrying". The author is careful to say these are lexical observations, not proof of a lack of progress. Besides /loop-report (add --agents for subagents and --history for closed cases), there are /loop-pane, which opens a live pane, and /loop-reset, which clears the observations. The README needs Claude Code 2.1.287 or later.

Use any of the methods in the install block; they install straight from the author's own marketplace. The status line shows loop: no active cases from the next session start.

The mod keeps a Detector object in memory. Every Bash call is given a fingerprint from the agent, the session directory and a hash of the exact command; the result is classified as failure, success or unknown. Here is the hook, from hooks/register.tsx:

let answer
try {
  // Never attach a fallback that could execute this tool a second time.
  answer = await next(e)
} catch (error) {
  if (ticket) safe(() => own.finish(ticket!, {}))
  if (own === detector) paint($)
  throw error
}

The tool runs first and its answer is returned as it came. Afterwards finish in hooks/core/detector.ts compares the failure signature with the previous one for that group. A different signature or a lower failure count restarts the streak. A success closes the case as recovered, a different failure as changed, and denied, interrupted, timed-out, background or oversized (over 65,536 characters) results as unknown. Idle cases expire after 200 later tool starts.

Two options are available in the plugin settings (/plugin menu):

The threshold is read when the mod loads or observations reset.

We read every file the plugin can execute before running anything: plugin.json, hooks/hooks.json, hooks/register.tsx and the four modules in hooks/core/ (detector.ts, fingerprint.ts, redact.ts, report.ts). They import only each other. There is no package.json, install script, binary, minified code, eval or dynamic import. The repository has an Apache-2.0 LICENSE file at its root.

Intercepts tool calls appears because tool.call is hooked without a matcher, so the mod sees every tool call, including the text of each Bash command. It only observes: it never denies or changes a call, and an error in its own code is swallowed so the tool still runs. The code uses $.session.cwd, $.command.register, $.ui.status, $.ui.toast, $.ui.open, $.ui.resolve, $.ui.invalidate and $.clock.after. It makes no network requests, runs no processes, writes no files, reads no environment variables or secrets, has no telemetry and starts no background service. Changes the UI covers the status line, toast and pane; Adds commands covers the three slash commands.

Only hashes, counters and bounded display text (160 characters) are kept in memory, with common secrets masked on a best-effort basis; nothing survives a reload, /clear or /loop-reset. claude plugin validate reported no errors or warnings, only notes that the tool.call and command.run hooks have no .catch; these hooks never block anything. The author's 41 tests passed on Claude Code 2.1.303.

Checked with Claude Code 2.1.303 on macOS (validate, tests and one interactive run). The mod's status text takes the place of the ? for shortcuts hint in the footer. If no case opens:

• The command must be byte-for-byte identical; a changed argument, quote or directory starts a new streak.
• The failure must be recognized: a nonzero exit code, an engine error or a known test-failure summary.
• Output that changes between runs, such as a timestamp, makes the failures different. Only some test-runner timings are normalized.
• Failures from subagents are tracked separately; use /loop-report --agents.
• VS Code chat and headless sessions have no status line or pane; use /loop-report.

To check that it works, ask Claude to run ls missing-dir three times. In our test run the status line changed to loop: 1 active · max 3 failures, a toast read "Loop Detective: case #1, 3 identical failures", and /loop-report then printed the case.