logslim

Extract signal from LLM session transcripts — decisions, errors, commands, files changed.

Turn 10MB session logs into 10KB summaries. Zero dependencies.

Install

npm install -g @aaaaorg/logslim

Usage

# Summarize a session (markdown output)
logslim session.jsonl

# Quick stats
logslim session.jsonl --stats
# be716422...  160 msgs  162 tools  10m  0 decisions  0 errors  32 files

# JSON output
logslim session.jsonl --json

# Plain text
logslim session.jsonl --format text

# Pipe from stdin
cat session.jsonl | logslim --stdin

# Multiple files
logslim *.jsonl --stats

# Strip tool call analysis
logslim session.jsonl --no-tools

# Unlimited content per message (default: 500 chars)
logslim session.jsonl --max-content 0

What it extracts

• Decisions — "chose to", "decided", "going with", "skipping"
• Errors — error messages from tool outputs
• Files touched — read/write/edit operations + paths in output
• Commands run — shell commands executed via exec tools
• Key exchanges — trimmed user/assistant conversation (first 3 + last 2)
• Stats — message count, tool calls, duration, model used

Output formats

Markdown (default)

# Session: be716422-...
**Started:** 2026-02-14 01:00 | **Duration:** 10m | **Messages:** 160 | **Tool calls:** 162

## Decisions
- Skipping codebase mapping...

## Errors
- Error: ENOENT no such file...

## Files Touched
- /path/to/file.js

Stats (one-liner per file)

be716422...  160 msgs  162 tools  10m  0 decisions  0 errors  32 files

JSON

Full structured object with all extracted data.

Input format

Reads JSONL files as produced by OpenClaw / Claude session logs. Each line is a JSON object with type, timestamp, and message fields.

Why

LLM session transcripts are huge (multi-MB) and mostly noise — tool call payloads, base64 images, repeated context. logslim extracts the 1% that matters: what was decided, what broke, what files changed.

Use cases:

• Knowledge extraction — feed summaries into memory systems instead of raw transcripts
• Cost auditing — quick stats across hundreds of sessions
• Debugging — find errors without reading 10MB of logs
• Archiving — keep summaries, delete raw logs

License

MIT