Native Launch Certification

Run the native agent launch-readiness gate with:

npm run test:native-launch-certification

The suite is intentionally offline and fixture-backed so it can run locally and in CI without live model spend. It certifies the launch surface for native OpenRouter models such as Kimi, Qwen, and GLM across the wavemill lifecycle.

Coverage

  • Native planning reaches planning/awaiting_user without auto-approval and preserves explicit approvals created after the awaiting-user state is visible.
  • Native coding applies scoped patches, writes completion artifacts, and supports blocked-completion handoff without writing .coding-complete.
  • Native review handles malformed model output, records review artifacts, and drives the PR handoff flow with idempotent GitHub/Linear fixtures.
  • Native launcher preflight rejects unsupported model/stage combinations before pane launch and emits actionable diagnostics.
  • Router certification verifies Kimi, Qwen, and GLM aliases and raw OpenRouter IDs against launch-priority role eligibility.
  • Lifecycle fixtures cover malformed native launchers, missing markers, monitor recovery, approval gates, blocked-completion behavior, and review handoff.
  • Dashboard/status fixtures verify native launch failures are surfaced to the operator instead of leaving panes stuck silently.

CI runs this command as a dedicated step after the normal shell/unit suite, so a native launch certification failure blocks launch-readiness changes.

Current Native Launch Matrix

This table reflects the checked-in native OpenRouter configuration and the phase preflight gate. It is the operator-facing runability matrix for native launches, not a general list of models that the router knows about.

Model alias OpenRouter model Planning Coding Review
glm-5.2 z-ai/glm-5.2 yes yes yes
kimi-k2.7-code moonshotai/kimi-k2.7-code yes yes yes
qwen-3-coder qwen/qwen3-coder no yes yes

Native coding is enabled by the checked-in nativeAgent.allowedPhases and nativeAgent.patchCoding.enabled settings. The repo-level patch-coding smoke artifact certifies the native patch runtime, and the provider/model workflow certification artifacts certify the configured OpenRouter model identities.

qwen-3-coder is review-executable but not planning-executable because its launch-priority role eligibility is coding and review, not planning.

Runability Versus Certification

Native model rollout has two separate gates:

  1. Certification and router eligibility prove that a model identity is known, has a valid native capability entry, and passes the fixture-backed native workflow certification suite.
  2. Launch runability proves that the exact phase/model pair can pass preflight for the current repository configuration and environment.

A model can be certified but still not launchable for a specific phase. The remaining qwen-3-coder planning rejection is the main example: the model is known and certified for native execution, but its role eligibility excludes planning.

Use this preflight command when checking a specific launch path:

node --import tsx tools/check-native-agent-launch.ts \
  --repo-dir . \
  --agent native-openrouter \
  --phase planning \
  --model kimi-k2.7-code

Change --phase to planning, coding, or review, and use either the model alias or raw OpenRouter model ID. A launch is executable only when this command passes in the same environment that will run wavemill mill.

Minimum Native OpenRouter Requirements

  • OPENROUTER_API_KEY is present in the launch environment.
  • nativeAgent.enabled is true.
  • nativeAgent.allowedPhases includes the phase being launched.
  • nativeAgent.providers.openrouter.enabled is true.
  • The raw OpenRouter model ID is listed in nativeAgent.providers.openrouter.models.
  • The model registry entry has a native capability for native-openrouter.
  • The model has a valid workflow certification artifact for the current certification suite.
  • Native coding additionally requires nativeAgent.patchCoding.enabled and a valid .wavemill/native-agent/patch-coding-certification.json smoke artifact.

Run the HOK-2074 canary fixture in tests/fixtures/native-routing-canary/README.md before broad rollout. It checks the current primary and challenger native model matrices against a real Wavemill routing task.