Skip to content

models: migrate legacy GPT routes and fallbacks to GPT-6 Luna - #3278

Merged
Hillary Mutisya (hillary-mutisya) merged 1 commit into
mainfrom
dev/hillarym/gpt6-luna-model-migration
Nov 11, 2026
Merged

Hillary Mutisya (hillary-mutisya) merged 1 commit into
mainfrom
dev/hillarym/gpt6-luna-model-migration

Conversation

@hillary-mutisya

Copy link
Copy Markdown
Contributor

Summary

  • Migrate explicit browser extraction, discovery, WebFlow, search, email, greeting, KnowPro, and document-processing selectors from GPT-3.5/4/5 to GPT-6 Luna.
  • Remove implicit older-model availability fallbacks for Azure chat defaults, dispatcher translation, and GPT-6-family Copilot requests.
  • Update provisioning defaults and sample configuration, and add regression coverage for routing, missing-Luna failures, and serialized KnowPro requests.

Motivation

The shared chat defaults already use GPT-6 Luna, but explicit legacy selectors and fallback chains could still route requests to older deployments. Browser helpers that construct Azure settings directly bypass provider-mode remapping, so changing only the shared default does not fix those calls.

Behavior and compatibility

  • Azure synthesizes its implicit bare chat endpoint only from the Luna deployment. Explicitly configured bare endpoints remain supported.
  • Dispatcher translation falls back only to Luna, preserves the configured reasoning effort, and reports a configuration error if neither the requested endpoint nor Luna is available.
  • An unavailable GPT-6-family Copilot request can fall back only to Luna, regardless of configured fallback lists. Missing or disabled Luna produces an unavailable-model error rather than silently selecting GPT-5 or another model family.
  • Explicitly requested, available models remain supported. This does not prohibit user-selected legacy models or change fallback behavior for explicitly configured non-GPT-6 Copilot requests.
  • KnowPro uses one accurately named Luna helper. The chat example exposes default/luna selections and preserves its older CLI aliases. Completion settings now enter the factory before normalization, avoiding a post-construction temperature override that Luna rejects.
  • Migrated Azure routes require a configured GPT_6_LUNA deployment; no deployment, secret, or local configuration is provisioned by this change.

Reviewer guide

  1. Review ts/packages/aiclient/src/copilotModels.ts and ts/packages/config/src/runtime/build.ts for availability/default policy, then the dispatcher fallback.
  2. Review browser/agent selectors and the self-host/provider-switch defaults for consistent routing.
  3. Review the KnowPro helper and example compatibility, followed by routing, configuration, provisioning, and request-body tests.

Validation

Validated locally on Windows:

  • pnpm run build: passed across the full workspace. A prior pnpm run clean followed by pnpm run build also passed across 328 tasks in 152 packages.
  • Nine selected Jest suites covering chat defaults, Copilot transport, API settings, endpoint pools, runtime configuration, provider mapping, endpoint availability, Azure configuration building, and dispatcher defaults: 172 passed.
  • pnpm run ci:node --test tools/scripts/test/knowledgeModelDefaults.spec.mjs tools/scripts/test/msiInstallerProviderUi.spec.mjs: 16 passed.
  • pnpm --filter @typeagent/browser run test:crossword: 71 passed.
  • pnpm run prettier:changed: passed.
  • All four committed-change gates passed against origin/main: lint ratchet; complexity ratchet at 25/30 with matching new-file limits; circular-dependency ratchet with the existing exception file; test-debt gate.

The local checkout also contains unrelated, uncommitted memory-curation work that is not included in this PR. No live-model inference, browser recording, quality campaign, or performance benchmark ran. Existing build/configuration warnings are not presented as failures or as fixed by this PR.

Risks and follow-ups

  • Installations relying on implicit older-model fallback must configure/enable Luna or explicitly select an available alternative.
  • The repository inventory still contains separate older-model selections in .NET benchmarks, service-key diagnostics, GitHub Models documentation workflows, workflow/benchmark examples, and ARM provisioning templates. These are not migrated here because they need their own provider/deployment validation.
  • Compatibility identifiers, pricing/capacity metadata, and historical test fixtures are deliberately retained; their names alone do not prove older-model execution.

Migrate explicit browser, email, greeting, KnowPro and document-processing selectors. Require Luna for implicit Azure defaults and unavailable GPT-6 Copilot/dispatcher routes while preserving explicit model overrides. Normalize KnowPro completion settings before construction and cover routing, missing-model failures and provisioning defaults.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@hillary-mutisya
Hillary Mutisya (hillary-mutisya) merged commit f3b26a7 into main Nov 11, 2026
111 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant