v0.3.0

10x OSS Scaler — Creativity, Throughput, Quality

Phase 1: Stabilize + Quick Wins

• Repair loop: up to 2 retries with error context injection after failed quality gates
• Test destruction guard: rejects repairs that destroy all tests
• 2+ builds per VDay (was 1)
• Decoupled demo agent: fire-and-forget, only runs when last build shipped
• Failure context injection in model cascade (previous tier errors inform next tier)
• Build metrics tracking via feedback-store.ts

Phase 2: Scale Throughput + Quality

• Parallel build slots (MAX_CONCURRENT=3, up to 5 builds/VDay)
• Smart work prioritization: feature items boosted +20, recently-failed repos deprioritized
• Test-driven repair: parse exact test errors, fix expected values not source
• Critic-to-builder feedback pipeline: structured hints from failure patterns
• Per-build budget cap ($0.30)

Phase 3: Creativity + Intelligence

• AI-powered scout feature mining: reads source code, suggests real features for high-quality repos
• Builder inspiration injection: recent successes + critic hints in prompts
• raceChat(): fire fast + standard in parallel, take first valid result
• Prompt evolution system: self-improving rules that reinforce on success, penalize on failure
• Self-improvement cycle: every 10 builds, analyzes trends, sends velocity reports

Targets

Metric	Before	After
Ships/day	1-2	12-15
Ship rate	65%	90%+
Feature ratio	20%	60-70%
Budget util	5%	36%

v0.2.0 — First Ship

v0.2.0 — First Successful Ship

The scaler achieved its first autonomous ship: fertigation-mix v3.0.1

Bug Fixes

• Compile gate: /dev/null as outdir fails on Linux — now uses /tmp. Tries multiple entry points
• Ship readiness: 94/100 builds were blocked — composite >= 80 now ships outright
• Scout re-queuing: Failed items re-queued infinitely — now deduped against completed work
• add-ci disabled: PAT lacks workflow scope — scout uses AI for real upgrades instead
• Stale state overwrite: Scout reloads fresh state before final save

Stats

• 42 API calls, $0.07 total spend on dedicated OR key
• 5 agents running end-to-end: Scout -> Builder -> Critic -> Demo -> Maintainer
• Slicing Pie leaderboard: Scout 21pts, Demo 8pts, Critic 8pts

v0.1.0 — Initial Release

OSSFactory-Scaler v0.1.0

Autonomous AI agent team for scaling OSS repositories.

5 Agents

• Scout — Scans repos, audits quality, prioritizes work queue
• Builder — Generates upgrades via 3-tier model cascade, pushes releases
• Demo — Creates SEO-optimized demo pages, deploys to VPS
• Maintainer — Triages issues, labels, health scoring
• Critic — Quality gates, daily observations

Features

• 4-tier OpenRouter model cascade (Gemini Flash Lite -> Flash -> Kimi K2.5 -> DeepSeek R1)
• 25 VDays/day on $5/day budget (~$0.20/VDay)
• Slicing Pie reward tracking with team leaderboard
• Quality gates: compile, tests, readme, security scanning
• Telegram notifications via oss_forge_bot
• Atomic JSON persistence, PM2 ready
• Zero runtime npm dependencies
• 16 tests passing