[Hillary Mutisya (hillary-mutisya)]

Summary

Add GPT-6.1 Terra deployment support and move active generation defaults to GPT-6/6.1, while retaining historical comparison models and availability fallbacks.

Azure configuration and model selection

• Preserve optional endpoint model metadata through typed config, YAML, legacy POOL environment projections, Azure settings, and request serialization. This lets generic /openai/v1/responses endpoints send the required deployment name in the request body without changing existing deployment-specific URLs.
• Document a gpt_6_1_terra Responses deployment with Sweden/East US capacities of 850/150 in the sample configuration. Configured deployments are discovered automatically by model selection.
• Export GPT-6.1 Sol/Terra selectors and provide Copilot and Ollama mappings.

Runtime and provisioning defaults

• Move active GPT-5.6 Sol requests to GPT-6.1 Sol, GPT-5 requests to GPT-6.1 Terra, and GPT-5 Nano/GPT-5.6 Luna defaults to GPT-6 Luna across agents and supporting tools.
• Use GPT-6 Luna for Foundry Bing grounding, URL resolution, phrase generation, and fast internet lookup.
• Keep Copilot's GPT-6 Luna default and extend same-tier model-availability fallback to GPT-6.1: an unavailable GPT-6/6.1 tier tries GPT-5.6 of the same tier before generic candidates.
• Extract default Azure endpoint selection without changing its behavior so the complexity ratchet continues to pass.

Benchmarks, corpus generation, and evaluation

• Default benchmark generation/review to GPT-6.1 Sol; use the GPT-6.1 Sol/Terra and GPT-6 Luna trio for ambiguity probes and collision-corpus generation.
• Move the action-picker default/environment override into its CLI instead of shell-specific parameter expansion.
• Pin new Copilot evaluation runs to GPT-6 Luna and reject reuse of older-model admission/accounting ledgers.
• Preserve explicit historical/comparison models, historical outputs, and existing pricing data. No unverified GPT-6 prices are added.

Deployment and fallback notes

• config.local.yaml is ignored and not included. The requested local Terra endpoints/capacities and local Copilot default were configured and validated separately; the PR ships the schema, sample, and code, not credentials or local environment provisioning.
• Foundry must separately have a gpt-6-luna deployment. The existing project's newly provisioned deployment was verified before testing Bing grounding; this PR does not provision shared Azure resources.
• Copilot fallback covers model availability, not arbitrary request failures. Azure pools, Foundry agents, and direct raw Copilot SDK callers do not acquire a universal GPT-5.6 fallback from this change.

Validation

• Full pnpm run build passed on Node v26.12.1.
• Config tests: 323 passed; AI-client tests: 475 passed across 30 suites.
• Focused grammar tests: 7 passed; benchmark config/ambiguity/action-picker tests: 27 passed; benchmark local suite passed.
• Provisioning/MSI/evaluation Node tests: 63 passed.
• Utility browser lifecycle: 7 passed; PowerShell analyzer: 4 passed; browser reasoning: 4 passed; onboarding: 31 passed.
• Changed-file formatting, whitespace, lint, complexity, circular-dependency, and test-debt ratchets passed against origin/main on committed changes.
• Live Azure Responses checks passed for Terra in both regions, GPT-6.1 Sol, and GPT-6 Luna.
• Live Copilot checks passed for the exact Sol/Terra/Luna models without fallback; Luna streaming and non-streaming both worked.
• Live Foundry Luna Bing-grounding smoke test returned an answer with a Bing citation. The isolated temporary agent was deleted afterward; existing agents were not modified.