scan

Scan — Generate Trigger Phrases

Overview

Scan one or more files to generate training-data-resonant trigger phrases. Each file is studied by 5 independent agents per target model; a research agent identifies convergent patterns to produce the most reliable semantic handle for each model.

Triggers are model-scoped — different models have different internal lookup keys. Specify which models your agents use so triggers are generated for the right audience.

Process

1. Parse Arguments

• File or directory path: What to scan
• --models <list>: Comma-separated list of target models (e.g., --models sonnet,opus). These are the models your agents run on. Default: sonnet
• No path: Ask the user what to scan

2. Resolve Input

• If the user provided a file path: use it directly
• If the user provided a directory: glob for relevant source files
  • Include: **/*.{ts,tsx,js,jsx,py,go,rs,md,json,yaml,yml,sh,sql,css,scss,html,svelte,vue}
  • Exclude: node_modules/, dist/, build/, .git/, *.lock, *.min.*
  • Cap at 20 files per run. If more, inform the user and process the first 20.
• If no argument: ask the user what to scan

3. Invoke Researcher Agent

Spawn the researcher agent (use the Agent tool with subagent_type set to total-recall:researcher). Provide it with the resolved file list and target models:

Generate trigger phrases for these files.
Target models: <comma-separated model list>

Files:
<list each absolute file path, one per line>

The researcher will spawn 5 study agents per model per file and perform per-model convergence analysis.

4. Store Results

Read .claude/triggers.json (create if it doesn't exist with {"version": 2, "triggers": {}}).

For each file result from the researcher, merge into the triggers object using the model-scoped format:

{
  "version": 2,
  "triggers": {
    "src/auth/middleware.ts": {
      "models": {
        "sonnet": {
          "phrase": "jwt route authentication guard",
          "samples": ["all", "five", "raw", "phrases", "..."],
          "convergence": ["jwt", "route", "auth"],
          "confidence": 0.90,
          "phases": 1
        },
        "opus": {
          "phrase": "jwt bearer token route guard",
          "samples": ["all", "five", "raw", "phrases", "..."],
          "convergence": ["jwt", "bearer", "route", "guard"],
          "confidence": 0.98,
          "phases": 1
        }
      },
      "crossModelTerms": ["jwt", "route"],
      "scannedAt": "2026-04-01T10:00:00.000Z"
    }
  }
}

Use relative paths (from project root) as keys. When adding new model results to a file that already has triggers for other models, merge into the existing models object — do not overwrite other models' triggers.

5. Rebuild Index

After storing results, invoke the total-recall:index skill to rebuild the master word index from all triggers.

6. Report

Display results as a table:

Target models: sonnet, opus

File                        | Model  | Trigger Phrase                 | Confidence
--------------------------- | ------ | ------------------------------ | ----------
src/auth/middleware.ts      | sonnet | jwt route authentication guard | 0.90
src/auth/middleware.ts      | opus   | jwt bearer token route guard   | 0.98
src/db/connection.ts        | sonnet | database connection pool mgmt  | 0.85
src/db/connection.ts        | opus   | database pool lifecycle rules  | 0.95

Cross-model terms: jwt, route, database, pool

If any file had low confidence (< 0.6), note it and suggest rescanning or model escalation.

Migration Note

If .claude/triggers.json has "version": 1 (old format with flat phrases), the scan will overwrite entries for scanned files with the new model-scoped format. Unscanned files retain the old format until re-scanned.