Summary

• Problem: Issue Codex migration (2026.6.1) drops the gpt-5.5 model when a canonical openai provider exists for embeddings — agents go silent #90047 reports that the 2026.6.1 doctor migration silently drops the entire openai-codex provider when any canonical openai provider already exists — including cases where both serve disjoint purposes (e.g. openai for embeddings via API key, openai-codex for chat via OAuth/ChatGPT backend). After a routine update, gpt-5.5 was gone and every agent turn failed.
• Root Cause: In migrateLegacyOpenAICodexProvider (legacy-config-migrations.runtime.models.ts:997), the else branch for the shadowed-provider case recorded "Removed" and called delete providers[providerId]. The already-normalized normalized.value — containing the API-renamed model entries — was discarded without checking for model rows absent from the canonical provider. The branch assumed a canonical openai provider is a superset of openai-codex, which is false when the two served disjoint models.
• Fix: Replace the silent drop with a targeted merge. For each model entry in the normalized legacy provider: (1) skip those whose id already appears in the canonical provider's model id Set or whose name appears in its name Set (separate Sets to avoid cross-type false positives); (2) stamp the legacy provider's baseUrl and api onto merged models that lack their own, so they continue routing to the correct endpoint (e.g. chatgpt.com/backend-api not api.openai.com/v1); (3) skip entries with neither id nor name to avoid duplicating malformed records. Only when the filtered set is empty is the old "Removed … already exists" message emitted.
• What changed:
  • src/commands/doctor/shared/legacy-config-migrations.runtime.models.ts — replace the else drop with a merge-or-remove branch; separate canonicalModelIds/canonicalModelNames Sets; stamp legacy provider baseUrl/api onto merged models that lack them.
  • src/commands/doctor/shared/legacy-config-migrate.test.ts — 3 new tests: disjoint merge with baseUrl preserved (the exact reporter scenario); partial-conflict merge (one duplicate skipped, one new entry with endpoint stamped); all-duplicate scenario still triggers the "Removed" message.
• What did NOT change (scope boundary):
  • The hasCanonicalOpenAIProvider check and normalizeLegacyOpenAIResponsesApi are unchanged.
  • The move branch (!hasCanonicalOpenAIProvider) is unchanged.
  • Config schema, defaults, and other doctor migrations are unchanged. Plugin surface unchanged.

Reproduction

1. Configure models.providers.openai (embeddings, API key, baseUrl: https://api.openai.com/v1) and models.providers.openai-codex (chat, OAuth, baseUrl: https://chatgpt.com/backend-api, models: [{ id: "gpt-5.5" }]).
2. Run openclaw doctor --fix or openclaw update on OpenClaw 2026.6.1.
3. Before this PR: migration deletes openai-codex entirely; gpt-5.5 is gone; agents fail.
4. After this PR: migration merges gpt-5.5 (api → openai-chatgpt-responses, baseUrl: https://chatgpt.com/backend-api stamped) into openai.models[]; both endpoints preserved; agents continue working.

Real behavior proof

Behavior addressed (#90047): migrateLegacyOpenAICodexProvider now merges disjoint model entries with correct endpoint metadata instead of dropping them.

Real environment tested (Ubuntu Server Node 24, source-level — Vitest against migrateLegacyConfigForTest with the reporter's exact multi-provider config): new tests reproduce the exact upgrade scenario and assert that the merged model array contains the correct baseUrl and renamed api.

Exact steps or command run after this patch: node scripts/run-vitest.mjs src/commands/doctor/shared/legacy-config-migrate.test.ts src/commands/doctor/shared/legacy-config-migrations.runtime.models.test.ts — 116 tests pass (113 pre-existing + 3 new). pnpm exec oxlint clean. pnpm format:check clean.

Evidence after fix:

Tests  116 passed (116)

New tests: merges disjoint model entries from legacy codex into canonical openai and preserves legacy baseUrl (#90047) — asserts openai.models = [{ id: "text-embedding-3-small" }, { id: "gpt-5.5", api: "openai-chatgpt-responses", baseUrl: "https://chatgpt.com/backend-api" }]; skips already-present model ids when merging legacy codex into canonical openai — one duplicate skipped, one new entry stamped with legacy baseUrl; removes openai-codex when all its models already exist in canonical openai — canonical unchanged, "Removed" message emitted.

What was not tested: live openclaw doctor --fix run against a real multi-provider config file.

Repro confirmation: the disjoint-merge test fails on the pre-patch tree (openai.models only contains the embeddings entry, gpt-5.5 absent) and passes with the fix.

Risk / Mitigation

• Risk: Wrong endpoint after merge. Mitigation: legacy provider baseUrl/api are stamped onto merged models that lack their own, tested by the new baseUrl assertion.
• Risk: Duplicate model rows. Mitigation: separate id/name Sets for dedup; cross-type false positives eliminated; all-duplicate path tested with "Removed" message assertion.
• Risk: Config surface change requiring Peter's approval. Mitigation: bug fix in an existing migration's behavior — no new config key, schema, or migration added.

Change Type (select all)

• Bug fix

Scope (select all touched areas)

• Doctor / migrations

Linked Issue/PR

Fixes #90047