agent

@workshop/agent

The constant core of the workshop — the code-review agent and its runtime, imported unchanged by all three patterns (naive, worker, workflow). Nothing here knows about Render; the substrate decides how the agent is invoked.

The review pipeline:

prepareDiff → filterDiff → [ security ‖ performance ‖ ux? ] → judge

Key exports

• runReview(prUrl, options) — the whole pipeline; returns a verdict, findings, and token usage. Emits progress via an onEvent callback (used by the worker).
• defineAgent(def) → an Agent with an in-process .run(input, ctx).
• The agents: securityReviewer, performanceReviewer, uxReviewer, judge, plus REVIEWERS, selectReviewers(), and hasFrontendFiles().
• prepareDiff(pr) — GitHub PR URL → per-file patches (public repos, no auth).
• filterDiff(patches) — drop noise files before review.
• parseDecision(text) — judge output → { verdict, reason, findings }.
• resolveClient / resolveModelSpec / MODEL_TIERS — provider-agnostic model client over fetch (Anthropic + OpenAI), with a deterministic mock client when no API key is set (or AGENT_MODEL=mock).
• Tools: defineTool, defineMcpSource, getToolRegistry, loadTools, registerTool, resolveTools — drop a file in src/tools/ to register.

Layout

src/
  review.ts        runReview + parseDecision (the orchestration)
  agents.ts        agent definitions + reviewer selection
  agent.ts         defineAgent() → in-process .run()
  loop.ts          the provider-agnostic LLM loop
  model.ts         Anthropic/OpenAI adapters + MockClient
  model-tiers.ts   small/medium/large → concrete models
  prepareDiff.ts   GitHub PR → Patch[]
  filterDiff.ts    noise filtering
  helpers.ts       shared utilities (review, patch, model tier)
  tool.ts / tool-registry.ts   tool helpers + auto-discovered registry
  tools/             one file per tool (auto-discovered by loader.ts)
  types.ts         shared contracts
  index.ts         public surface

No API key needed — the mock model lets the full pipeline run offline.