0.1.8 Beta5: Run the full maintained DeepSeek V4 Flash 0731 Responses regression #393

Parent

#351 (part of #248)

Outcome

Run the full maintained-Provider regression for the Ollama Cloud deepseek-v4-flash:0731 Responses route against every generic Responses capability shipped in Beta2 through Beta4.

This is independent maintenance evidence for a bundled Provider configuration. It does not inherit GLM-5.2 or K2.7 Code results, select production behavior, or restrict user-configured endpoints.

Required coverage

• Beta2: exact route identity; native/adapt/omit/required-unavailable behavior; standard functions; structural mapping; hosted-tool omission; V1/V2 isolation; no cross-Provider proxy.
• Beta3: complete client-executed tool_search, reversible Code Mode, and reasoning/stream classifications.
• Beta4: complete client-owned Collaboration V2 lifecycle and user-owned agent configuration safety.
• Retained Beta1 and Beta3.1–3.3 continuation, history, context-window/output-limit, visibility, catalog ownership, and identity regressions.

Acceptance criteria

• Streaming/non-streaming text, multi-turn history, standard function calls/results/replay, and stable terminal/error behavior complete through the real Codex CLI on the Ollama Responses endpoint.
• Every relevant Beta2 structural disposition is exercised.
• Client tool_search, Code Mode, and Collaboration V2 complete through the same generic production implementation.
• V2 task identity, message/agent_message, follow-up, wait/result, list/interrupt, streaming/history, and same-Home restart/readback are covered.
• Optional unsupported hosted tools are absent; a required unavailable capability fails before sampling.
• Exact Provider/model identity remains deepseek-v4-flash:0731; no Official/other-model fallback or cross-Provider execution occurs.
• Slow-valid streams, upstream availability, incomplete SSE, client cancellation, and protocol incompatibility remain distinct.
• A failure is not fixed by adding a DeepSeek-specific codec or routing branch.
• Evidence is bounded and bound to exact candidate SHA, CLI version, Provider/model/protocol, and catalog identity.

Out of scope

• Chat Completions, owned by Beta6.
• A runtime model allowlist or claim that arbitrary custom endpoints are qualified.

Blocked by

• 0.1.8 Beta4 gate: Publish the runtime-derived generic Collaboration V2 candidate #284