retrolens 0.5.4

Debugger-style AI conversation log navigator — let any agent extract lessons and workflows from chat sessions

RetroLens 🔬

A debugger for AI conversations — let any agent navigate, analyze, and learn from past chat sessions.

What is this?

Every day you use AI agents (VS Code Copilot, Claude Code, Cursor…) to write code. Those conversation logs are full of reusable workflows and hard-won lessons — but they're locked inside opaque log files.

RetroLens is a lightweight CLI + Skill that gives any general-purpose AI agent the ability to:

1. 📂 Discover log sessions across platforms
2. 🔍 Navigate sessions like a debugger — overview → turn → tool call
3. 💡 Analyze what happened and extract lessons, workflows, and agent directives

The CLI handles only what agents can't do well — parsing complex log formats and providing a structured traversal API. All analysis and writing is done by the agent itself, guided by SKILL.md.

┌─────────────────────────────────────────────┐
│  AI Agent (Copilot / Claude / Cursor / ...) │
│                                             │
│  Reads SKILL.md → learns how to use CLI     │
│  Calls CLI → gets structured JSON data      │
│  Analyzes → uses its own LLM reasoning      │
│  Writes → LESSONS.md, AGENTS.md, etc.       │
└──────────────────┬──────────────────────────┘
                   │ CLI calls
                   ▼
┌─────────────────────────────────────────────┐
│  retrolens CLI                              │
│                                             │
│  cfg  → configure log path & source         │
│  ls   → list sessions                       │
│  read → drill into turns & tool calls       │
└──────────────────┬──────────────────────────┘
                   │ Parses log files
                   ▼
┌─────────────────────────────────────────────┐
│  Log Files                                  │
│  VS Code Copilot · Claude Code · Custom     │
└─────────────────────────────────────────────┘

Integrating this skill

retrolens must be installed for agent use.

pip install retrolens
retrolens --version

Then add the skill to your agent.

Claude Code

This repo ships a Claude marketplace that installs the retrolens-plugin/ bundle as a plugin.

Step 1: Add marketplace

/plugin marketplace add JoelYYoung/retrolens

Step 2: Install the plugin

/plugin install retrolens

Manual install alternative:

git clone https://github.com/JoelYYoung/retrolens.git
cp -r retrolens/retrolens-plugin ~/.claude/plugins/retrolens
/reload-plugins

VS Code Copilot

VS Code Copilot discovers skills from your repo at .github/skills/<name>/SKILL.md.

From this repo, copy the skill into your target project:

mkdir -p .github/skills/retrolens
cp retrolens-plugin/SKILL.md .github/skills/retrolens/SKILL.md

Generic Skill-Based Agents

If your agent supports local skill folders, copy the skill file:

cp retrolens-plugin/SKILL.md <your-agent-skills-dir>/retrolens/SKILL.md

Then prompt the agent to follow the workflows in SKILL.md while calling the retrolens CLI.

Quick Start

You can specify what to analyze and what artifact to produce.

Common workflows:

1) Find sessions for this project

Use the RetroLens skill to locate conversation logs for this repository and list the most recent 10 sessions (title, date, turn count).

2) Add support for a new log format (custom reader)

This is only needed if your logs are not from VS Code Copilot or Claude Code.

Use the RetroLens skill: logs for a new platform are in <LOG_DIR>. Sample the files, implement a custom reader, and validate it can parse turns and tool calls correctly.

3) Analyze a specific session (ID/prefix: XXX)

Use the RetroLens skill to analyze session XXX: summarize the goal, key pivots, failures and fixes, then write reusable directives to notes/session-XXX-lessons.md.

4) Postmortem the latest session

Use the RetroLens skill to postmortem the latest session and produce 5 human-facing lessons and 5 agent directives in notes/retrolens-latest.md.

Providing the project path, log directory (or platform: VS Code / Claude Code), and the desired output file name will speed things up.

---

How It Works — the Skill Model

RetroLens is designed to be used by AI agents, not by humans directly. The included SKILL.md teaches any general-purpose agent how to navigate and analyze conversation logs through a series of workflows:

Workflow	What the Agent Does
Discover & Connect	Find log directories, auto-detect format, verify with ls
Build Custom Reader	Implement a reader for unsupported log formats
Analyze a Session	Progressive drill-down: ls → read → read --turn → read --tool, then map workflow phases
Reflect & Extract Lessons	Categorize findings into human lessons and agent directives
Cross-Session Mining	Compare multiple sessions for recurring patterns

The agent reads structured JSON from the CLI and does all reasoning, categorization, and writing itself.

---

CLI Reference

cfg — Configure Working State

retrolens cfg set --path <dir>        # Set log directory (auto-detects format)
retrolens cfg set --source vscode     # Override detected source type
retrolens cfg set --reader ./r.py     # Register a custom reader
retrolens cfg show                    # Show current config
retrolens cfg clear                   # Reset

ls — List Sessions

retrolens ls                          # List sessions (default: 20)
retrolens ls --limit 50 --json        # JSON output, up to 50

read — Navigate Session Data

retrolens read <ID>                   # Session overview
retrolens read <ID> --turn N          # Turn N detail
retrolens read <ID> -t N --tool M     # Tool call M in turn N
retrolens read <ID> --turns 1-5       # Range of turn summaries
retrolens read <ID> --diff 1,3        # Compare two turns
retrolens read <ID> --raw -t N        # Raw JSON data for a turn
retrolens read <ID> --json            # JSON output (for agents)

<ID> can be a full UUID, a prefix (e.g. fb48c), or latest.

Supported Formats

Platform	Source Type	Format
VS Code Copilot Chat	vscode	JSONL (incremental state)
Claude Code	claude_code	JSONL (event stream)
Custom	any	Via custom reader (see SKILL.md Workflow B)

Custom Readers

For unsupported log formats, create a Python file that subclasses BaseReader:

from pathlib import Path
from retrolens.readers import BaseReader
from retrolens import models

class MyReader(BaseReader):
    source_type = "myplatform"

    def scan(self, path: Path | None = None) -> list[models.SessionInfo]: ...
    def get_overview(self, session_id: str) -> models.SessionOverview: ...
    def get_turn(self, session_id: str, turn_number: int) -> models.TurnDetail: ...

Register it: retrolens cfg set --reader ./my_reader.py

Full guide: retrolens-plugin/SKILL.md → Workflow B, or retrolens-plugin/references/READER-API.md.

Bundled Scripts

Script	Purpose
retrolens-plugin/scripts/setup.sh	Install RetroLens and verify
retrolens-plugin/scripts/sample_log.py	Sample & pretty-print a log file to identify format
retrolens-plugin/scripts/validate_reader.py	Validate a reader against real log data

Project Structure

src/retrolens/
├── cli.py                  # Click CLI (cfg, ls, read)
├── config.py               # Persistent config state
├── detect.py               # Log format auto-detection
├── models.py               # Pydantic v2 data models
├── formatters.py           # Text / JSON dual-mode output
└── readers/
    ├── __init__.py         # BaseReader ABC + ReaderRegistry
    ├── vscode_copilot.py   # VS Code Copilot JSONL parser
    └── claude_code.py      # Claude Code JSONL parser
retrolens-plugin/
├── SKILL.md                # ⭐ Agent skill document
├── scripts/                # Helper scripts (setup, discovery, validation)
└── references/             # Reader API docs
tests/                      # 128 tests

Development

uv pip install -e ".[dev]"       # Install with dev deps
python -m pytest tests/ -v       # Run tests
ruff check src/                  # Lint

License

MIT — see LICENSE