Ecosystem Overview

The wasmagent ecosystem spans three repositories that work together as a single product:

┌────────────────────────────────────────────────────────────────┐
│                        wasmagent-js                            │
│  Agent Runtime / SDK / Kernel / Ranking / Model Adapters       │
│  • Secure WASM sandboxing (QuickJS, Pyodide, Wasmtime)         │
│  • 8+ model adapters incl. Chinese providers                   │
│  • RolloutForkRunner → training data hooks                     │
└──────────────────────┬─────────────────────┬───────────────────┘
                       │ npm packages        │ rollout JSONL
          ┌────────────▼──────────┐  ┌───────▼────────────────────┐
          │       bscode          │  │        trace-pipeline       │
          │  Edge-native coding   │  │  Model merge / eval /       │
          │  agent template       │  │  RLAIF data factory         │
          │  • Cloudflare Worker  │  │  • TrainingDataExporter     │
          │  • React/Next.js UI   │  │  • DPO/PPO generation       │
          │  • build + visual     │  │  • Eval harness             │
          │    verification       │  │  • trace-pipeline optimizer │
          └───────────────────────┘  └────────────────────────────┘

The three layers

Layer	Repository	User value
Runtime	wasmagent-js	Secure portable agent runtime: WASM isolation, tool governance, model adapters, observability, rollout ranking
Reference App	bscode	One-click deployable coding agent on Cloudflare Workers — proves the runtime works end-to-end
Data Factory	trace-pipeline	Converts real agent runs into DPO/PPO training data, closes the improvement loop

How the loop closes

bscode runs jobs
  → build result (pass/fail) + visual checks
  → wasmagent-js RolloutForkRunner records trajectories
  → AEPEmitter emits signed AEP evidence record (aep/v0.2, Ed25519)
  → RolloutRanker scores branches (objective + judge)
  → rollout-wire JSONL + AEP bundle (Layer 1 + evidence)
  → trace-pipeline validate-aep → trust-score → audit-report
  → trace-pipeline TrainingDataExporter
  → DPO/PPO training records (Layer 3)
  → model fine-tune / merge / eval
  → improved model back into bscode defaults

The AEP (Agent Evidence Protocol) is the cross-repo public data contract: @wasmagent/aep emits records at runtime; trace-pipeline validate-aep and audit-report consume them before training export. This makes every training record traceable to a specific agent run with a verified trust score.

This loop is what separates wasmagent from a generic agent framework — it produces training signal from real deployments, not synthetic benchmarks.

Where to start

• Using the runtime → Getting started with wasmagent-js
• Deploying bscode → bscode README
• Running the data pipeline → Data Pipeline guide
• Schema contract between repos → Schema Governance
• Integration smoke test → trace-pipeline/tests/test_three_repo_smoke.py

Repository links

Repository	Description	Docs
wasmagent-js	Agent runtime, npm packages	This site
bscode	Cloudflare coding agent	bscode/README.md
trace-pipeline	Model merge + RLAIF factory	trace-pipeline/docs/