[jtaroreh]

Why

Operators dispatching Cloud Run swarms faced high initial friction when selecting models:

1. Gated partner models (Anthropic Claude) require manual Google Cloud Console agreement forms that autonomous agents cannot accept programmatically.
2. Open-weights models (GLM 5.3, Llama 3.3) require dedicated 8x H100 GPU cluster endpoints rather than serverless scale-to-zero compute.
3. Preflight probes lacked automatic Vertex AI fallback, alias resolution for DeepSeek MaaS, and generationConfig limits, which caused preflight latency and burnt token quota.

This change equips Cloud Run swarms with zero-friction Tier 1 serverless models (Gemini 4 Argon and DeepSeek V4.1 Flash) on global endpoints.

Scope

• Categorized Cloud Swarm models into Tier 1 (Native MaaS serverless), Tier 2 (Gated Partner MaaS), and Tier 3 (Dedicated GPU endpoints).
• Added --list-models to cloud_dispatch.py to inspect supported tiers and enablement requirements directly from CLI.
• Added preflight interception in cloud_dispatch.py for Tier 3 open-weights models (glm-5-3-flash, llama-3.3-70b) with actionable diagnostic guidance.
• Implemented automatic Vertex AI promotion and endpoint routing for DeepSeek MaaS under the global namespace (publishers/deepseek-ai/models/deepseek-v4.1-flash-maas).
• Restored generationConfig: {"maxOutputTokens": 1} across Vertex AI generateContent probe calls to eliminate preflight token waste.
• Decoupled Anthropic SDK check from check_model_garden() in setup_runtime.py so Tier 1 serverless operators pass doctor checks without optional partner dependencies.
• Added comprehensive unit tests in test_cloud_preflight_and_worker.py and test_cloud_worker_vertex_partner.py.

Tradeoffs

• Dedicated GPU endpoints (GLM 5.3, Llama 3.3) fail-closed during preflight with explicit remediation guidance rather than attempting invalid Cloud Run dispatches.
• DeepSeek MaaS routes strictly through the Vertex AI global location namespace, which is the sole supported region for DeepSeek MaaS endpoints on Google Cloud.

Blast Radius

• skills/swarm/scripts/cloud_dispatch.py: Affects preflight probe logic and swarm model alias resolution. Backward compatible with all existing Gemini and Claude partner models.
• skills/swarm/scripts/cloud_worker.py: Resolves per-task model overrides and sets global location for DeepSeek tasks.
• skills/setup-agystack/scripts/setup_runtime.py: Broadens Model Garden detection to include native serverless models.

Verification

• Adversarial review conducted across 3 subagents (pro, flash, inherit). See report in artifact adversarial_review.md.
• Full unit test suite passes: 315 tests run, 315 passed, 0 failures (python3 -m unittest discover tests).