claudbg/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

claudbg is a Rust CLI tool for inspecting Claude Code sessions. It reads JSONL session files from ~/.claude/projects/, caches them in a local libsql database (~/.claude/claudbg.db), and exposes commands to list, dump, and transcribe sessions and sub-agent runs.

Development Environment

The project uses Nix flakes with crane for building. Enter the dev shell before working:

nix develop

The dev shell provides: rustToolchain (stable + rust-analyzer + clippy), cargo-watch, cargo-edit, cargo-nextest, clang, beans, jq.

Use jq for all JSON processing in shell commands — do not use python3.

Common Commands

# Build
cargo build

# Run
cargo run -- <args>

# Tests (preferred runner)
cargo nextest run

# Single test
cargo nextest run <test_name>

# Watch mode (rebuild on changes)
cargo watch -x build
cargo watch -x 'nextest run'

# Lint (deny warnings — same as CI)
cargo clippy --all-targets -- --deny warnings

# Format check
cargo fmt --check

# Format fix
cargo fmt

# Run all checks via Nix (matches CI)
nix flake check

Planned Architecture

The v0.1.0 milestone targets a working session inspector with these layers, in dependency order:

1. CLI Foundation (src/main.rs, src/cli/)

• clap (derive) wires top-level subcommands: sessions, agents, index, tui, query
• Global flags on every command: --output [table|json|xml] (default: table), --verbose, --include <thinking,output,tools> (default: tools)
• AppError enum + Result<T> alias using thiserror
• Short UUID display utility (8-char prefix)

2. Data Parsing & Models (src/models/)

• Rust types (serde-deserializable) for the full JSONL schema:
  • SessionEntry enum: user / assistant / system / progress / file-history-snapshot / queue-operation
  • AssistantContent enum: text / thinking / tool_use
  • UserContent enum: text / tool_result / image
  • Usage, ToolUseInput, SystemMessage subtypes
• Session file discovery: walk ~/.claude/projects/, URL-decode directory names to recover project paths, return SessionRef { session_id, project_path, file_path, modified_at }
• Sub-agent file discovery (parallel structure)
• JSONL line-by-line reader
• SessionStats computation: token counts, tool call breakdown by name, duration, model

3. Database & Caching (src/db/)

• libsql (local file, async via tokio) at ~/.claude/claudbg.db
• Normalized tables: sessions, messages, tool_uses — used for queries and list views
• Denormalized table: raw_sessions (session_id PK, raw_jsonl TEXT, file_mtime, file_size) — used for fast dump
• Lazy sync: on any read command, check file_mtime/file_size against DB; re-parse and upsert if stale
• index subcommand: force full resync of all discovered session files
• --clear flag: drops and recreates the DB

4. Output Formatting (src/output/)

• Trait-based renderers behind the --output flag
• Table (default): comfy-table (in use — render_table in src/output/table.rs)
• JSON: serde_json pretty-print
• XML: hand-rolled or quick-xml

5. Sessions Commands (src/commands/sessions/)

• sessions list: sorted most-recent-first; columns: short session ID, date, project path, model, message count; --verbose shows full UUID
• sessions dump <id>: table rows — #, timestamp, type, role, content preview (truncated); --verbose shows full content; accepts 8-char prefix or full UUID
• sessions transcribe <id>: header with SessionStats, then chat log (user text, assistant text/tool-use blocks, thinking blocks gated by --include thinking, tool output gated by --include output)
• --follow flag: watch the session file and stream new entries as they arrive

6. Agents Commands (src/commands/agents/)

• Mirror of sessions commands but for sub-agent JSONL files
• agents list <session-id>, agents dump <session-id> <agent-id>, agents transcribe, --follow
• Note: discover_agents_for_session(session_file) must look in session_file.stem()/subagents/, not session_file.parent()/subagents/

7. Stub Commands

• tui: prints "coming soon!"
• query: prints "coming soon!"

Key Design Decisions

• Session IDs are UUIDs; display as 8-char prefixes by default, full with --verbose
• ~/.claude/projects/ directory names are URL-encoded project paths (e.g. -run-media-pop-... → /run/media/pop/...)
• The DB acts as a read-through cache; the JSONL files on disk are the source of truth
• All async runtime: tokio (full)
• Dependency crate choices (from beans tickets): clap (derive), tokio (full), serde + serde_json, libsql, chrono, uuid, thiserror, comfy-table or tabled

JSONL Schema Notes

• Field names in Claude Code JSONL are camelCase — serde fields need #[serde(rename = "...")]:
  • durationMs (not duration_ms) — on top-level system entries
  • sessionId (not session_id) — on all entries
  • parentUuid, promptId, gitBranch, etc.
• Sub-agent files live at <project-dir>/<session-uuid>/subagents/agent-<id>.jsonl — the session UUID is also a directory alongside <session-uuid>.jsonl; subagents/ is inside that directory, not next to the .jsonl

Issue Tracking

This project uses beans for issue tracking. Always use beans instead of TodoWrite for task management. See beans --help for usage.