Universal Audit Log

Mods observability 21 Scanned by NVIDIA SkillSpector

Install Command

npx claude-code-templates@latest --mod observability/universal-audit-log

View on GitHub

claude-code — universal-audit-log

Scanned by NVIDIA SkillSpector

Automated static analysis (no LLM) · v2.12.0 · 2026-10-17

Partial analysis: some checks could not run on this component.

2 observations

scanner score 29/100

SkillSpector looks for patterns such as shell commands, credential access, network calls or unpinned dependencies. An observation is something worth reading in the files before you install, not a confirmed vulnerability. Components that run code locally, like hooks, MCP servers and security tooling, naturally show more of them.

Agent Config Directory Access
Agent Snooping · README.md:30
MCP server referenced without pinned version: 'npx claude-code-templates'.
MCP Rug Pull · README.md:39

Learn more about NVIDIA SkillSpector.

Content

Claude Mod. A mod is a complete Claude Code plugin built on function hooks (Anthropic's mods/ layout); it is on by default in Claude Code 2.1.287+. Hooks module: hooks/universal-audit-log.ts. Options: path, skipEvents, flushEvery, maxBytes (set in user or managed settings pluginConfigs, or /config). Run claude plugin validate on the installed folder to see every event it hooks and every $ call it makes.

universal-audit-log

One hook on "" sees every event: the engine's own (tool.call, prompt.submit, turn., ...) and every other plugin's calls on $ (fs.read, http.fetch, process.run, ...). It records one JSON line per dispatch: the event, who raised it (next.origin: plugin and tier), how long it took and whether something beneath denied or threw. The "after" placement, so the outcome is recorded too; denials surface here as the value next(e) resolves to.

Lines are buffered in memory and flushed to a JSONL file through $.fs (there is no append on $, so a flush reads the file, keeps its tail under maxBytes, and writes it back) at the end of every turn and every flushEvery events. The hook is never re-entered for its own $.fs calls, so flushing from inside it does not recurse.

Seat this plugin FIRST (an org prepends it in managed settings) so nothing beneath it can bypass the log. A hook that fails is skipped by the engine (fail-open, logged in --debug-file), so it never blocks the session.

Options

  path:       string    JSONL file, relative to the working directory (default ".claude/logs/mods-audit.jsonl")
  skipEvents: string    events not to record, comma-separated (default: the render / log chatter)
  flushEvery: number    flush after this many buffered lines (default 50)
  maxBytes:   number    keep the file under this size, dropping the oldest lines (default 2 MiB)

Declared in .claude-plugin/plugin.json (userConfig). Set them in /config, in user settings (~/.claude/settings.json, not project settings), with --settings <file> or in managed settings:

json

{ "pluginConfigs": { "universal-audit-log@skills-dir": { "options": { } } } }

Install

npx claude-code-templates@latest --mod observability/universal-audit-log
claude

It is written to .claude/skills/universal-audit-log/, which Claude Code auto-loads as universal-audit-log@skills-dir. For one session with hot reload: claude --plugin-dir .claude/skills/universal-audit-log. claude plugin validate .claude/skills/universal-audit-log prints every event it hooks and every $ call it makes.

Requirements. Mods are on by default in Claude Code 2.1.287+. Typed against Anthropic's declarations: https://github.com/anthropics/claude-code/tree/main/mods