Skip to content

SESSION_LOGGING.md

Latest commit

 

History

History
89 lines (62 loc) · 2.76 KB

SESSION_LOGGING.md

File metadata and controls

89 lines (62 loc) · 2.76 KB

Session Logging System

This system uses dual logging - two complementary approaches for different purposes.

The Two Logs

Log Type Location Created By Purpose
Summary Logs Claude Dashboard/Session Logs/ Claude (per CLAUDE.md) Quick reference, searchable summaries
Raw Transcripts LLM Logs/Claude Code Logs/ raw_log.js script Full conversation history, debugging

Summary Logs

What they are: Condensed records of each exchange.

Format:

## 14:30
**User:** Asked to refactor the authentication module
**Actions:**
- Read `src/auth/login.py`
- Edit `src/auth/login.py` - extracted validate_token function
- Write `src/auth/tokens.py` - new token utilities module
**Response:** Refactored auth module, extracted token handling to separate file

---

How they're created: Claude follows CLAUDE.md instructions to log after each exchange.

When to use them:

  • "What did we do yesterday?"
  • "Which file did we edit for that bug fix?"
  • Quick session overview

Raw Transcripts

What they are: Full conversation exports with all tool calls, outputs, and thinking.

Format: Markdown with collapsible sections for:

  • Tool calls and their outputs
  • Claude's thinking blocks
  • System messages

How they're created: Manual export using raw_log.js after sessions.

When to use them:

  • Debugging what went wrong
  • Reviewing exact command outputs
  • Training/improving prompts
  • Full context recovery

Important: Neither Auto-Fires

Summary Logs: Claude does this during the conversation, but only because CLAUDE.md instructs it to. If Claude forgets or you don't have CLAUDE.md set up, no logs.

Raw Transcripts: Requires manual export. The hook system doesn't auto-trigger on session end (known limitation).

Sync Feature

The raw_log.js script includes automatic sync:

  • When you export, it also syncs any missing sessions
  • Scans all JSONL files in your Claude projects directory
  • Compares against exported files in vault
  • Exports any that are missing

Result: Run the export once, get all your sessions. No gaps accumulate.

# First run might sync many:
Synced 85 missing session(s) to vault

# Subsequent runs:
(no output = already synced)

Workflow

  1. During session: Claude logs summaries per CLAUDE.md
  2. After session: Run claude-export alias to save raw transcript (also syncs any missed sessions)
  3. Next session: Claude reads previous session log for context

Tips

  • Export raw logs every ~5 exchanges or at natural breakpoints
  • Don't rely solely on raw logs - they're verbose and slow to search
  • Summary logs are your quick reference; raw logs are your backup
  • Both are plain markdown - searchable in Obsidian