[jacobprall]

Fixes #67. Replaces #70 with one change.

The Cursor SDK (#68) and Pydantic AI (#69) adapters are no longer in this PR. They're kept on the harness-cursor and harness-pydantic branches, still on the earlier runtime interface. Each ships once it runs as a driver on the kernel below.

An agent now names a harness: a value an adapter in app/harness/ builds from its library's own configuration, instead of a model-tier string. The model is whatever you gave the library:

pi({ model: getBuiltinModel("anthropic", "claude-opus-5") }) // typed: a mistyped model fails typecheck

config.models, ModelTier, and modelFor are gone. The built-in choices live in app/agents/models.ts as piSonnet and piOpus, each named for its library and model, and the defaults are unchanged: Triage and PR Reviewer on Sonnet, the spec writer and implementer on Opus.

The run kernel

• The kernel owns each run (app/runner/core/kernel.ts):
  • It emits started, then exactly one result or error, and picks the exit code.
  • It enforces max_turns, and timeout_ms minus its safety margin, and handles cancellation.
  • It builds the agent's factory tools and submit.
• The first outcome is final. That outcome is an accepted submission, the last rejected submission, the turn limit, the timeout, or cancellation. The driver then gets 5 seconds to stop before the outcome is reported anyway.
• A driver only translates its library (app/runner/core/driver.ts):
  • open(ctx) reports the library version and the exact model.
  • run() offers the model ctx.tools and ctx.submit, and calls ctx.requestModel() before every model request.
  • It reports what the library observed through ctx.report, and stops when ctx.signal aborts.
  • A driver never sees the emitter, the submission, or exit codes. The kernel drops a tool result that was never announced, and anything reported after the outcome.
• max_turns counts model requests. pi's requests pass the kernel's check in its streamFn, so retries of a transient model error count, and a retry past the limit is never sent.
• No hidden reminder turn. A model that stops without calling submit ends with no_submission.
• Production code has no test mode. Tests drive the kernel with pi's fake provider and a scripted driver with no library at all.

Runner and Sandbox

• Runner layout:
  • core/: the versioned contract (protocol 2), kernel, driver interface, emitter, submit checks, input parsing, and keys.
  • tools/: the shared factory tools, with a library-neutral view.
  • pi/: the pi driver and its tool bindings.
• The bundle: agent definitions are bundled into the runner, which runs the agent named in its input on that agent's harness. runner.mjs --self-test proves the bundle and every agent definition load.
• Payload: when a Sandbox opens and its runner bundle's hash differs from this deploy's, the bundle is swapped in, root-owned, before verify.sh. Harness, model, and agent changes need no snapshot refresh.

Also

• Credentials: keys come from the harnesses. runAgent, doctor, and the software-factory-workflows env group derive them.
• Telemetry: agent.finished records the harness, its version, and the model.
• Emitter fix: createEmitter's terminated getter was copied as a value by Object.assign, so it always read false. It now reflects the stream.

Verification

• npm run check: 167 tests pass, including:
  • 13 kernel unit tests on outcomes, limits, cancellation, the stop grace, tools, and reporting;
  • 5 pi driver tests on no reminder, retries counting as turns, a batch that submits, and an unknown model; and
  • the runner conformance suite on pi and the scripted driver.
• The tests catch a real bug: loosening the kernel's turn limit by one fails both the kernel and pi tests.
• On a test deployment, with this commit plus a demo-target config change:
  • One autopilot run on the kernel (jacobprall/software-factory-demo#17) ran all four agents on pi. Triage, the spec writer, the implementer, and PR Reviewer each ended ok, in 2, 2, 4, and 2 model requests, for about $0.13 in total. It opened draft PR #18 with exactly the requested file and posted a COMMENT review. The Workflows service logged no errors.
  • The run's first Sandbox swapped in the new runner bundle, with no snapshot refresh.
  • Timeouts, the turn limit, and rejected submissions are covered only by the tests above.
  • Earlier revisions on the same deployment covered pi-only Triage, a stale-payload swap, and a fresh-snapshot open.

Deploying

Deploy factory-workflows. Existing snapshots get the new runner bundle when each Sandbox opens.