satwork — Bounty Factory

GitHub Repo → Scan → Generate → Deploy → Optimize → PR Merged

Quick Start

# Scan a Go repo for optimization candidates
bounty-factory scan https://github.com/lightningnetwork/lnd

# Generate eval scripts + target configs from the scan
bounty-factory generate scan-report.json -o lnd-targets/

# Deploy targets to the satwork coordinator
bounty-factory deploy lnd-targets/ --coordinator https://satwork.ai

# After workers find improvements, generate a PR
bounty-factory pr lnd-ban-management --repo https://github.com/lightningnetwork/lnd

Supported Languages

Language	What it scans	Tested on
Go	Named constants (`Default`, `Min`, `Max*`), duration constants, `TODO:tune` comments, channel buffers, retry loops, benchmarks	lightningnetwork/lnd — 244 candidates, 40 benchmarks
Python	Module constants (`UPPER_SNAKE`), class defaults, function parameter defaults, config files, `TODO:tune`, pytest-benchmark	psf/requests — 5 candidates, 339 tests
Rust	`const`/`static` declarations, `Default` trait implementations, `TODO:tune`, criterion benchmarks, `#[bench]`	tokio-rs/tokio — 32 candidates, 44 benchmarks
JavaScript / TypeScript	`const` declarations, config objects, `.env.example` defaults, `TODO:tune`, benchmark suites	webpack/webpack — 4 candidates, 3,395 tests

How It Works

Phase 1: Scan

bounty-factory scan <github-url> clones the repo, detects the language, and runs two passes:

Pass 1: Named constants in const blocks, Default* prefixed values, duration constants, TODO:tune comments
Pass 2: Buried magic numbers — channel buffers, retry loop counts, slice pre-allocations, hardcoded thresholds in function bodies

Each candidate is semantically classified (timeout? buffer? probability? threshold?) and grouped into logical optimization targets. Output: scan-report.json.

Phase 2: Generate

bounty-factory generate <report> creates a complete satwork target for each group:

eval.py — Deterministic simulation matching the domain. Uses LCG PRNG, 3 seeds, median score, multi-metric composite scoring.
proposed_params.json — Current default values as the baseline
registration.json — Target config with parameter bounds, budget, reward
source_map.json — Maps each parameter back to file:line for PR generation

Phase 3: Deploy

bounty-factory deploy <targets-dir> pushes targets to the coordinator and hot-reloads. Workers discover new targets automatically via /api/discover.

Phase 4: PR

bounty-factory pr <target-id> --repo <url> fetches the winning parameters, maps them back to source code, and generates a PR with annotated diffs, benchmark comparisons, and verification instructions.

Semantic Classification

The scanner classifies each parameter by its variable name and surrounding context:

Category	Detected by	Bounds strategy
timeout	timeout, deadline, expiry, ttl	0.1x to 10x current value
buffer_size	buffer, page_size, batch_size, pool	0.25x to 4x
probability	prob, weight, factor, ratio, decay	[0, 1]
retry	retry, retries, max_attempts, backoff	±50%
concurrency	workers, threads, conns, goroutines	0.25x to 4x
interval	interval, period, tick, flush, rotation	0.1x to 10x
cache	cache_size, evict, lru, prealloc	0.25x to 4x
threshold	threshold, limit, max_, min_, cap	±50%
fee	fee, cost, price, rate_ppm, budget	±50%

Eval Templates

Generated eval scripts use domain-specific simulation templates:

Domain	Triggered by	Simulates
routing	routing/, htlcswitch/, autopilot/	Synthetic LN graph, payment routing, success rate + fees
database	channeldb/, sqldb/, graph/db/	Query patterns, batch commits, cache hit rates
gossip	discovery/	Message propagation, bandwidth, peer banning
connection	peer/, server, watchtower/	Peer connections, reconnection, keepalive
cache	cache/	Hit rates, eviction, memory efficiency
sweep	sweep/, contractcourt/	Fee environment, UTXO batching, deadline management
generic	everything else	Multi-metric sensitivity analysis

Live Results

Active targets

Live on the coordinator right now

Total proposals

Evaluated by distributed workers

Improvements found

Verified, reproducible, ready for PRs

Try it now. The Bounty Factory scanned lightningnetwork/lnd and generated live targets. Workers have found improvements of +4% to +19.8%. See the Lightning Labs presentation for the full demo.

Example: Scan Output

$ bounty-factory scan https://github.com/lightningnetwork/lnd

Cloning https://github.com/lightningnetwork/lnd...
  Language: go
  Commit: db765a752b63
  Found 308 candidates, 40 benchmarks, 1733 tests
  After confidence filter (>=0.3): 244 candidates
  Grouped into 107 targets

============================================================
SCAN SUMMARY: https://github.com/lightningnetwork/lnd
============================================================
  Candidates: 244
  Benchmarks: 40
  Tests: 1733
  Target groups: 107

  lnd-lncfg-timeout                         10 params  timeout
  lnd-htlcswitch-timeout                     7 params  timeout
  lnd-sqldb-buffer_size                      6 params  buffer_size
  lnd-routing-probability                    5 params  probability
  lnd-discovery-threshold                    7 params  threshold       [TODO:tune x3]
  lnd-graph-db-cache                         3 params  cache
  ...

Example: Generated PR

$ bounty-factory pr lnd-ban-management --repo https://github.com/lightningnetwork/lnd

PR GENERATED: perf: optimize LND Gossip Ban Management (TODO: tune) (+19.8%)
  Branch: bounty-factory/optimize-lnd-ban-management
  Changes: 4 parameters across 1 files

--- a/discovery/ban.go
+++ b/discovery/ban.go
- DefaultBanThreshold = 100
+ DefaultBanThreshold = 57

- maxBannedPeers = 10_000
+ maxBannedPeers = 1_000

- banTime = time.Hour * 48
+ banTime = time.Hour * 149

- resetDelta = time.Hour * 48
+ resetDelta = time.Hour * 57