Wayfinder: OpenAI GPT 5.6 (Sol/Terra/Luna) via Azure → Codeq (staging end-to-end)

[Kenny-Heitritter]

Destination

Sol / Terra / Luna (OpenAI GPT 5.6) selectable and working in Codeq on staging, served through Azure via the qBraid gateway (account-staging.qbraid.com/api/ai/v1). Prod Azure inference is provisioned (models deployed at a much-higher rate limit) but not user-facing — the prod cutover to Codeq users is a follow-on effort. DONE = pick a model in staging Codeq and get a real Azure-served completion, billed correctly.

Notes

• This map carries execution. The destination is a working change (staging-first), so task tickets do the work (terraform applies, PRs, deploys), not just decide. (Overrides wayfinder's plan-only default.)
• Four-layer stack across four repos: (1) Azure terraform — qbraid-infrastructure; (2) gateway — qbraid-account /api/ai/v1; (3) billing — qbraid-api; (4) Codeq client — opencode branding/qbraid/models.json.
• Decisions locked while charting: gateway built on qbraid-account (current home; accept rework if feat/remove-ai-gateway-from-account lands); Azure CI/CD wiring in scope; model identity needs discovery (root ticket).
• Related effort — map Wayfinder: CodeQ deployment — seamless staging/prod updates without stale binaries/configs #16 (CodeQ deployment, issues Execution: publish the CODEQ_DEPLOYMENT.md anchor doc to dev #23–29): the mechanism that gets a Codeq client change onto running staging pods. The models.json rollout should ride Wayfinder: CodeQ deployment — seamless staging/prod updates without stale binaries/configs #16's staging path (Execution: enable staging auto-rebuild (codeq push → build-lab-base, scoped IAM) #28), not duplicate it.
• Region: staging AI Foundry account qbraid-staging-ai-services is in eastus2; prod RG qbraid-prod likewise. Existing staging deployments follow the azurerm_cognitive_deployment gpt-5.x pattern in terraform/environments/staging/azure/.
• Staging Azure applies are unguarded. The destructive-change guard only runs for auto_apply: false environments (.github/actions/iac-destructive-guard/action.yml:63-65), so nothing gates a destroy on staging-azure. terraform/environments/staging/azure also does not currently plan clean — see Reconcile staging/azure terraform drift to a zero-diff plan. Any ticket that applies staging Azure terraform must read that plan before merging.
• Skills per session: /grilling, /domain-modeling, /research, /prototype.

Decisions so far

• Discover the real Azure identifiers + capabilities for Sol/Terra/Luna — Sol/Terra/Luna = 3 tiers of one generation (GPT-5.6) → 3 separate Azure deployments: gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna, all model_version 2026-07-09, GlobalStandard, eastus2 ✅ (doc-confirmed). Caps (all 3): ctx 1,050,000 / out 128,000, reasoning + tools + vision, temperature:false. Pricing $/1M: Sol 5/30, Terra 2.5/15, Luna 1/6 (cache read −90%, write ×1.25; >272K input = 2×). Prod ceiling = subscription quota tier (Tier 6 = 15M TPM; sku_capacity=TPM/1000) — live tier check spun to #41, which now blocks Bootstrap prod Azure inference + deploy the 3 models at a high rate limit #34. Schema gaps flagged for Add billing pricing + quota for the 3 models (qbraid-api) #38 (MODEL_PRICING lacks azure/cache/long-context) & Add Sol/Terra/Luna to the Codeq model catalog (branding/qbraid/models.json) #39 (context_over_272k).

Not yet specified

• Staging rollout of the Codeq client change to running pods — how the updated branding/qbraid/models.json actually reaches staging Codeq pods. Hangs on map Wayfinder: CodeQ deployment — seamless staging/prod updates without stale binaries/configs #16's staging auto-rebuild path (Execution: enable staging auto-rebuild (codeq push → build-lab-base, scoped IAM) #28) and on the models.json ticket being built. Graduates into its own ticket (which will also block the end-to-end verification) once Wayfinder: CodeQ deployment — seamless staging/prod updates without stale binaries/configs #16's staging path settles.

Out of scope

• Prod user-facing cutover — routing prod gateway traffic to Azure prod, prod creds plumbing, prod billing entries, prod rollout to Codeq users. Beyond the staging-first destination; a follow-on effort. (Prod Azure provisioning is in scope; exposing it to users is not.)
• Gateway migration to qbraid-api (feat/remove-ai-gateway-from-account) — we build on qbraid-account instead. Separate effort; carries rework risk if it lands first.
• Gateway-layer rate-limiting / APIM spend-limiter overhaul — this effort's "rate limit" is the Azure sku_capacity on the prod deployments (in scope). Per-user gateway limits / spend-limiter are separate.