For AI agents: the complete documentation index is available at https://scriptedalchemy.github.io/agent-bundle/llms.txt, the full documentation bundle is available at https://scriptedalchemy.github.io/agent-bundle/llms-full.txt, and this page is available as Markdown at https://scriptedalchemy.github.io/agent-bundle/guide/development/index.md.
agent-bundle
  • English

      • The development loop

      Development has one entry point and three evidence surfaces. The entry point is agent-bundle dev — a foreground, loopback-only server that rebuilds the artifact as inputs change. The surfaces are the developer Workbench, the test harness, and the eval runner.

      npx agent-bundle dev --root .

      The rebuild loop

      dev prepares project source, builds an artifact, and publishes it as an epoch: an immutable generation identified by an epochId. Every surface that reads a built artifact — the Workbench artifact tree, an MCP session, a hook simulation, a development host install — names the epoch it is reading, so a rebuild mid-session can never silently change what a result was evidence about.

      Rebuilds are debounced and serialized. A successful build publishes an artifact.available event; a failed build publishes nothing, so the previously published epoch stays exactly as it was. There is no partially published generation.

      The same serialized pass also rebuilds the dist/ package build when the project declares bin/lib entries. That build has its own provenance-based incremental boundary: after a success, the sorted source inputs of every emitted file are kept, and the next rebuild is skipped unless one of those inputs was invalidated, the configuration file, package.json, or tsconfig.json changed, the normalized bin/lib declaration or the tools hatch changed, the invalidation was manual or initial, or the previous package build failed. A brand-new file that changes module resolution without touching a tracked input is picked up on the next tracked change rather than instantly.

      A package build failure never invalidates the artifact epoch that already committed. It surfaces as one AB7103 warning on the succeeded build attempt and retries on the next invalidation.

      Development also publishes generated route declarations at .agent-bundle/routes.d.ts from the same compiled graph. Each write goes to a sibling temporary file and is renamed over the prior complete declaration atomically, so invalid source keeps the last-good file, and a successful route-free preparation removes it.

      The three surfaces

      SurfaceWhat it isWhat it proves
      Developer WorkbenchThe loopback UI dev serves: diagnostics, Skills, artifact provenance, MCP and hook playgrounds, Playground traces, eval runs.Nothing by itself — it is where you look at and exercise real generated output.
      Testingagent-bundle/rstest and agent-bundle/test, plus the framework-owned contract matrix.Recorded proof levels, from a route module rendering to a bundle spawned from an installed host layout.
      EvaluationsTyped eval suites run through deterministic, Claude, or Codex harnesses.Whether an agent actually reaches your plugin, with pass / fail / inconclusive and a declared minimum evidence bar.

      A plugin that builds is not a plugin that works, and each of these answers a different question. None of them substitutes for another.

      Where this section stops

      Authoring surfaces — what a Skill, hook, MCP route, script, or package entry is — are in Authoring. Turning a validated artifact into something a host can install is in Distribution. Exact flags, config field semantics, and runtime contracts are in Reference.