plog

PromptLog

Run PromptLog from this plugin package. The user does not need a global plog installation.

CLI runner

Use this exact runner for every CLI operation, preserving the current project as the working directory:

node "${CLAUDE_PLUGIN_ROOT}/dist/index.js"

Never substitute a global plog command and never install PromptLog from inside this skill. If the runner fails because Node is older than 22.13, report the detected Node version and the minimum requirement.

Arguments: $ARGUMENTS

Route the request

• No arguments: run status when .promptlog/config.json exists; otherwise run init, followed by the semantic audit below.
• init: run init, then perform the semantic audit when intelligence.mode is host-agent.
• scan: run scan --json, then perform the semantic audit when intelligence.mode is host-agent.
• audit: ensure the project is initialized, then perform only the semantic audit.
• watch: start the bundled watch command as a background process. Report how to stop it.
• ui: start the bundled ui command as a background process and report the localhost URL. Do not expose it on 0.0.0.0 unless the user explicitly requests that.
• Any other PromptLog subcommand: pass the arguments to the bundled runner unchanged. Supported commands include status, inventory, diff, note, notes, note-delete, export, rollback, add, ignore, unignore, and config.
• A plain-language phrase: map it to the closest action, state the action selected, and continue.

Do not leave foreground watch or ui processes blocking the session.

Semantic audit

The deterministic scanner is the first pass. This audit is an independent, repository-aware second pass designed to find prompt surfaces that naming and syntax heuristics can miss.

1. Read .promptlog/config.json. Respect scanner.include, scanner.exclude, and intelligence settings. If intelligence.mode is none, skip this audit and say so.
2. Run inventory --json. This inventory intentionally contains locations and hashes but not prompt text.
3. Locate actual model invocation boundaries in the included source: provider SDK calls, agent/framework constructors, message arrays, prompt templates, completion/chat/generation calls, and project-specific model wrappers.
4. Trace prompt-bearing arguments backward through helpers, imports, object properties, arrays, joins, template functions, and configuration. Also inspect instruction files that are loaded into an agent or model at runtime.
5. Treat a value as a prompt when its text or assembled result is supplied to a model as instructions, system/user/assistant content, a template, examples, tool guidance, evaluation criteria, or agent policy.
6. Exclude UI copy, logs, ordinary documentation, tests/fixtures, generated output, dependencies, and historical .promptlog data unless the application actually sends that content to a model at runtime.
7. Compare every confirmed prompt surface with the inventory by normalized source path and overlapping line range. Do not call something missed merely because its stable name differs.
8. Report four concise groups: tracked, missed, uncertain, and excluded false positives. For each missed or uncertain item, provide source file, exact line range, symbol or purpose, and reasoning. Do not print full prompt content when sendFullPromptByDefault is false.
9. Never register a missed prompt silently. When requireUserConfirmation is true, present the proposed add operations and wait for explicit approval. After approval, run one bundled add <file> --start <n> --end <n> --name <stable-name> operation per accepted item, then rerun inventory --json and summarize the result.