py-tt-debug 0.7.0

Python Time-Travel Debugger — record, replay, and step backward through program execution

pyttd

pyttd (Python Time-Travel Debugger) is an open-source time-travel debugger for Python with full VSCode integration. It records complete program execution at the C level, then lets you step backward and forward, jump to any point in the trace, and visually scrub through a timeline — all from within your editor.

What is Time-Travel Debugging?

Traditional debuggers only move forward. If you step past a bug, you start over. pyttd records every execution frame during a single run, then drops you into a replay session where you can navigate freely in both directions:

• Step backward through execution to see exactly how state evolved
• Reverse continue to find the last time a breakpoint was hit
• Jump to any frame in the entire recording
• Scrub a visual timeline to navigate through call depth, exceptions, and execution flow
• Inspect variables at any point without re-running the program
• Track variable history across the entire recording
• Set conditional, function, data, and log breakpoints — all work in both directions
• Attach to running processes to start recording mid-execution
• Export traces to Perfetto for external analysis

pyttd is a post-mortem replay debugger — your script runs to completion, and then you debug the recorded trace.

Installation

Requires Python 3.12+ — uses CPython C API features introduced in 3.12.

pip install py-tt-debug

Or from source:

git clone https://github.com/cydonknight/pyttd.git
cd pyttd
python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"

Quick Start

CLI

# Record a script
pyttd record examples/hello.py

# Record with arguments
pyttd record examples/hello.py --args --verbose --count 5

# Record only specific functions
pyttd record examples/hello.py --include process_data --include validate

# Query the recording
pyttd query --last-run --frames

# List all runs in a database
pyttd query --list-runs --db hello.pyttd.db

# Replay and jump to a specific frame
pyttd replay --last-run --goto-frame 750

# Export to Perfetto trace format
pyttd export --db hello.pyttd.db -o trace.json
# Open trace.json at https://ui.perfetto.dev

# Clean up database files
pyttd clean --all --dry-run

VSCode

1. Build the extension: cd vscode-pyttd && npm install && npm run package
2. In VSCode: Extensions sidebar → ... → Install from VSIX → select pyttd-*.vsix
3. Add to .vscode/launch.json:

{
    "type": "pyttd",
    "request": "launch",
    "name": "Time-Travel Debug",
    "program": "${file}"
}

4. Press F5 — your script records, then you navigate freely

To replay an existing recording without re-running, use an attach configuration:

{
    "type": "pyttd",
    "request": "attach",
    "name": "Replay Trace",
    "traceDb": "${workspaceFolder}/script.pyttd.db"
}

Python API

from pyttd import ttdbg

@ttdbg
def my_function():
    x = 42
    return x * 2

my_function()  # Records to <file>.pyttd.db

Or with explicit start/stop control:

from pyttd import start_recording, stop_recording

start_recording(db_path="trace.pyttd.db")
# ... your code ...
stats = stop_recording()

Attach to a Running Process

from pyttd import arm, disarm

# Start recording mid-execution
arm()
suspect_function()
stats = disarm()  # Returns recording stats

# Or use as a context manager
with arm():
    suspect_function()

# Or toggle via signal (kill -USR1 <pid>)
from pyttd import install_signal_handler
install_signal_handler()  # First USR1 arms, second disarms

Features

Recording

• C extension recorder using PEP 523 frame eval hooks (not sys.settrace)
• Lock-free per-thread SPSC ring buffers with background flush
• Fork-based checkpointing for fast cold navigation (Linux/macOS)
• Multi-thread recording with globally ordered sequence numbers
• Async/await and generator support with coroutine frame tracking
• I/O hooks for deterministic checkpoint replay — time.time, time.monotonic, time.perf_counter, time.sleep, random.random, random.randint, random.uniform, random.gauss, random.choice, random.sample, random.shuffle, os.urandom, uuid.uuid4, uuid.uuid1, datetime.datetime.now, datetime.datetime.utcnow
• Secrets filtering — sensitive variable names (password, token, secret, api_key, etc.) automatically redacted during recording; configurable patterns with --secret-patterns
• Selective function recording — --include / --exclude filter by function name, --include-file / --exclude-file filter by source file (glob patterns)
• Expandable variable trees — dicts, lists, tuples, sets, and objects with __dict__ are serialized with structure, not just repr()
• Runtime attach — arm() / disarm() API to start/stop recording from within a running process, or toggle via Unix signal
• Checkpoint memory tracking and configurable limits

Navigation

• Forward: step into/over/out, continue with breakpoints
• Reverse: step back, reverse continue with breakpoints and exception filters
• Conditional breakpoints — expressions evaluated against frame locals
• Hit-count breakpoints — stop after N hits (supports >=N, >N, <=N, <N, ==N, %N)
• Log points — emit structured log messages with variable interpolation without stopping
• Function breakpoints — break on any call to a named function
• Data breakpoints — break when a variable's value changes
• Jump: goto frame, goto targets (all executions of a line), restart frame
• Warm navigation (SQLite, sub-ms) for stepping; cold navigation (checkpoint restore, 50-300ms) for jumps

VSCode Extension

• Full DAP implementation with step-back and reverse-continue
• Canvas-based timeline scrubber with click/drag/zoom
• CodeLens annotations showing call and exception counts per function
• Inline variable values during stepping
• Call history tree with lazy-loaded nesting and exception markers
• Exception breakpoint filters (uncaught, all raised)
• Function breakpoints, data breakpoints, conditional breakpoints, hit conditions, log points
• Variable history tracking across the recording
• Breakpoint verification — validates that breakpoints target executable lines
• Status bar with recording progress (frame count, dropped frames, DB size)
• Keyboard shortcuts: Ctrl+Shift+F11 (step back), Ctrl+Shift+F5 (reverse continue)

Analysis & Export

• Perfetto/Chrome Trace Event Format export — viewable in ui.perfetto.dev; preserves multi-thread structure
• Variable history queries — track how a variable changes over time, with deduplication
• Execution stats — per-function call counts, exception counts, and entry points
• Checkpoint memory profiling — per-checkpoint RSS tracking

Database Management

• Multi-run storage — multiple recording runs in a single database
• Run eviction — --keep-runs N auto-evicts old runs; pyttd clean --keep N for manual cleanup
• Custom DB paths — --db-path overrides the default <script>.pyttd.db location
• Size monitoring — --max-db-size warns when the database exceeds a threshold
• Run selection — --run-id to query, replay, or export a specific run by UUID or prefix

CLI Reference

pyttd [--version] [-v|--verbose] <command>

Commands:
  record    Record script execution
  query     Query trace data
  replay    Replay a recorded session
  serve     Start JSON-RPC debug server (used by VSCode)
  export    Export trace data
  clean     Clean up database files

record

pyttd record script.py [options]

Options:
  --module                          Treat argument as a module name (e.g., pkg.mod)
  --checkpoint-interval N           Frames between checkpoints (default: 1000)
  --args VALUE [VALUE ...]          Arguments to pass to the script
  --no-redact                       Disable secrets redaction
  --secret-patterns PAT             Additional pattern to redact (repeatable)
  --include FUNC                    Only record functions matching this pattern (repeatable)
  --include-file GLOB               Only record functions in files matching this glob (repeatable)
  --exclude FUNC                    Exclude functions matching this pattern (repeatable)
  --exclude-file GLOB               Exclude files matching this glob (repeatable)
  --max-frames N                    Stop recording after N frames (0 = unlimited)
  --db-path PATH                    Custom database path (default: <script>.pyttd.db)
  --max-db-size MB                  Warn when DB exceeds this size in MB (0 = unlimited)
  --keep-runs N                     Keep only last N runs, evict older (0 = keep all)
  --checkpoint-memory-limit MB      Checkpoint memory limit in MB (0 = unlimited)

query

pyttd query [--last-run | --run-id UUID] [--list-runs] [--frames] [--limit N] [--db path.pyttd.db]

replay

pyttd replay [--last-run | --run-id UUID] --goto-frame N [--db path.pyttd.db]

serve

# Record and serve (used by VSCode extension)
pyttd serve --script script.py [--module] [--cwd DIR] [--checkpoint-interval N]
    [--include FUNC] [--exclude FUNC] [--include-file GLOB] [--exclude-file GLOB]
    [--max-frames N] [--env KEY=VALUE ...] [--env-file .env]
    [--db-path PATH] [--max-db-size MB] [--keep-runs N]

# Replay existing recording (no re-recording)
pyttd serve --db path.pyttd.db [--run-id UUID]

export

pyttd export --format perfetto --db path.pyttd.db [--run-id UUID] -o trace.json

clean

pyttd clean [--db path.pyttd.db] [--all] [--keep N] [--dry-run]

Options:
  --db PATH      Specific database file to clean
  --all          Delete all .pyttd.db files in current directory
  --keep N       Keep last N runs, evict the rest
  --dry-run      Show what would be deleted without deleting

Environment Variables

Variable	Description
PYTTD_RECORDING	Set to 1 during active recording; scripts can check os.environ.get('PYTTD_RECORDING')
PYTTD_ARM_SIGNAL	Auto-install signal handler on import — e.g., PYTTD_ARM_SIGNAL=USR1 installs a SIGUSR1 toggle handler

Architecture

Three-layer system:

Layer	Technology	Responsibility
C Extension (pyttd_native)	C, CPython API	Frame recording, ring buffer, checkpoints, I/O hooks
Python Backend (pyttd/)	Python, Peewee, SQLite	JSON-RPC server, session navigation, query API
VSCode Extension (vscode-pyttd/)	TypeScript	DAP handlers, timeline webview, CodeLens, inline values

See docs/architecture.md for the full design.

Platform Support

Platform	Recording	Warm Navigation	Cold Navigation	Multi-Thread
Linux	Full	Full	Full	Full
macOS	Full	Full	Partial*	Full
Windows	Full	Full	None	Full

* macOS: checkpoints skip when multiple threads are active.

Requirements

• Python >= 3.12 (required for PyUnstable_InterpreterFrame_* C API)
• C compiler (GCC, Clang, or MSVC)
• VSCode (for the extension; CLI works standalone)

Known Limitations

• Expression evaluation operates on recorded snapshots, not live values
• C extension internals are opaque (third-party C extension objects may have uninformative repr())
• Windows: no cold navigation (no fork())
• exception_unwind line number is from function entry, not the exception site
• Variable repr strings are capped at 256 characters
• Expandable variable children are capped at 50 entries, 1 level deep
• Attach mode (arm()) disables checkpoints — cold navigation is unavailable for attached recordings
• Tight loops with per-line events have high overhead (hundreds of times slower); use --include to scope recording for compute-heavy code

Testing

351 Python tests across 26 test modules + 95 VSCode extension (Mocha) tests:

# Run all Python tests
.venv/bin/pytest tests/ -v

# Run VSCode extension tests
cd vscode-pyttd && npm test

# Run overhead benchmarks
.venv/bin/python3 benchmarks/bench_overhead.py

Benchmarks cover I/O-bound, compute-bound, tight loop, deep recursion, many locals, and multi-thread workloads.

Documentation

• Getting Started — first recording walkthrough
• CLI Reference — all commands and flags
• VSCode Guide — extension features and configuration
• API Reference — Python programmatic API
• Architecture — system design and data flow
• Troubleshooting — common issues
• FAQ — frequently asked questions
• Contributing — how to contribute
• Changelog — version history

Development guides: Building | Testing | C Extension | Protocol | Releasing

Contributing

Contributions welcome across C, Python, and TypeScript. See CONTRIBUTING.md for setup and guidelines.

License

MIT License. See LICENSE for details.