Session Transcript Analyzer

Session Transcript Analyzer

Step 1: Locate the transcript

Determine the target session. Two modes:

• Explicit session ID: pass --session <uuid> to scan ~/.claude/projects/<project-dir>/ for a .jsonl file whose name contains the UUID.
• Auto-detect (no --session): scan the project directory (derived from process.cwd()) for the most recently modified .jsonl file. Fall back to a global scan of ~/.claude/projects/ if the project directory does not exist.

Project folder name derivation: replace the drive colon+backslash with --, then replace remaining path separators with -. Example: C:\dev\projects\agent-studio → C--dev-projects-agent-studio.

Step 2: Pair with a debug log

Once the session ID is extracted from the transcript filename (<uuid>.jsonl), look for a matching debug log at ~/.claude/debug/<uuid>.txt. If not found, do a prefix/contains scan of the debug directory. Warn but continue if no debug log is found.

Step 3: Parse the JSONL transcript

Read the file line by line. Skip blank or malformed lines. For each valid JSON object:

• Count user messages: event.type === 'user' where event.message.content[0].type === 'text'
• Count tool invocations: content blocks with type === 'tool_use'; accumulate per-tool counts in a frequency map
• Collect tool failures: content blocks with type === 'tool_result' and is_error === true

Step 4: Parse the debug log

Read line by line. Collect lines that match any of:

• [ERROR] or error: — general errors
• BLOCKED — hook blocks (router lockdown, creator guard)
• prompt is too long — context-length API errors
• SyntaxError: Unterminated string in JSON — hook JSON parse failures

Step 5: Run heuristics and generate report

Build a Markdown report with these sections:

1. Overview Metrics: user turns, tool invocations, tool errors, debug log error count
2. Tool Usage Summary: sorted frequency table of tool names
3. Heuristic Findings (flag each category if count > 0):
   • Router Protocol Violations (ROUTER-LOCKDOWN, TASK-CREATE VIOLATION)
   • Context Length Exceeded (prompt is too long)
   • Hook JSON Parse Failures (Unterminated string in JSON)
   • If none triggered: print a clean-system message
4. Top Tool Failures: first 5 tool_result errors with truncated content (150 chars)

Step 6: Write report

Default output path: .tmp/transcript-analysis-<first-8-chars-of-uuid>.md (create .tmp/ if needed). If --out <path> was supplied, write there instead.

1. Tolerate malformed lines: wrap JSON.parse() in try/catch per line; never abort on a single bad line.
2. Warn, don't fail, on missing debug log: the transcript alone provides useful metrics; missing debug log loses only heuristic signal.
3. Truncate tool failure content to 150 chars to keep the report readable; collapse newlines to spaces.
4. Derive project dir from process.cwd() not from a hardcoded path — supports running from any agent-studio checkout location.
5. Show only the first 5 tool failures in the report; full data is in the transcript for deeper inspection.

node scripts/analyze-session-transcript.mjs
# or
pnpm debug:analyze

Analyze a specific session by UUID:

node scripts/analyze-session-transcript.mjs --session f1326443-490f-486b-bb67-01c72bf42408

Write report to a custom path:

node scripts/analyze-session-transcript.mjs --out .claude/context/reports/backend/session-analysis-2026-03-10.md

Typical report output structure:

# Session Analysis Report: f1326443-490f-486b-bb67-01c72bf42408

Generated: 2026-03-10T12:00:00.000Z

## Overview Metrics

- **User Turns:** 14
- **Tool Invocations:** 87
- **Tool Errors:** 3
- **Debug Log Errors:** 2

## Tool Usage Summary

- `Read`: 34 times
- `Edit`: 18 times
- `Bash`: 15 times
- `Write`: 12 times
- `TaskUpdate`: 8 times

## Heuristic Findings

### Context Length Exceeded (API Error)

The orchestrator accumulated too many tokens and crashed the API request.
Occurred **1** times.

## Top Tool Failures (from Transcript UI)

- `toolu_01ABC`: File not found: .claude/context/runtime/stale-tasks.json...

Search Protocol

For code discovery and search tasks, follow this priority order:

1. `pnpm search:code ""` (Primary intent-based search).
2. `ripgrep` (for exact keyword/regex matches).
3. semantic/structural search via code tools if available.

Memory Protocol (MANDATORY)

Before starting: ```bash cat .claude/context/memory/learnings.md cat .claude/context/memory/decisions.md ```

After completing:

• New pattern -> `.claude/context/memory/learnings.md`
• Issue found -> `.claude/context/memory/issues.md`
• Decision made -> `.claude/context/memory/decisions.md`

ASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.