Summary

• Problem: Ollama dynamic discovery and dynamic model preparation overwrote an explicit provider API like openai-completions with api: "ollama", sending requests through the native /api/chat path instead of the configured OpenAI-compatible route.
• Solution: Keep native Ollama base URL normalization for discovery probes, but preserve the user's explicitly configured provider API and base URL on the final provider/runtime model when the API is non-native.
• What changed: Updated the Ollama discovery merge and dynamic model conversion paths, with regression tests for catalog discovery, explicit configured models, native Ollama normalization, and prepareDynamicModel/resolveDynamicModel.
• What did NOT change: Native Ollama discovery probing, default api: "ollama" behavior, other providers, core routing, or plugin SDK contracts.

Fixes #93710

Real behavior proof

• Behavior addressed: Explicit models.providers.ollama.api: "openai-completions" and baseUrl: ".../v1" remain configured after Ollama discovery even though the live builder uses native api: "ollama" and a native base URL for /api/tags probing.
• Real environment tested: macOS 26.5, Node v24.15.0, OpenClaw 2026.6.2, worktree SHA 67cb1a0
• Exact steps or command run after this patch:

  node --import tsx --input-type=module -e 'import { resolveOllamaDiscoveryResult } from "./extensions/ollama/src/discovery-shared.ts"; const result = await resolveOllamaDiscoveryResult({ ctx: { config: { models: { providers: { ollama: { baseUrl: "http://192.168.20.14:11434/v1", api: "openai-completions", models: [] } } } }, env: { OLLAMA_API_KEY: "test-key" }, resolveProviderApiKey: () => ({ apiKey: "test-key" }) }, pluginConfig: {}, buildProvider: async () => ({ baseUrl: "http://192.168.20.14:11434", api: "ollama", models: [{ id: "qwen3-coder:cloud", name: "qwen3-coder:cloud" }] }) }); const provider = result?.provider; console.log(JSON.stringify({ configuredApi: "openai-completions", configuredBaseUrl: "http://192.168.20.14:11434/v1", discoveredBuilderApi: "ollama", discoveredBuilderBaseUrl: "http://192.168.20.14:11434", resolvedProviderApi: provider?.api, resolvedBaseUrl: provider?.baseUrl, resolvedModelIds: provider?.models?.map((model) => model.id) ?? [] }, null, 2));'
  node scripts/run-vitest.mjs extensions/ollama/provider-discovery.test.ts
  node scripts/run-vitest.mjs extensions/ollama/index.test.ts
  pnpm check:test-types
  node scripts/run-oxlint.mjs --deny=warn --ignore-path=.oxlintignore extensions/ollama/index.test.ts extensions/ollama/index.ts extensions/ollama/provider-discovery.test.ts extensions/ollama/src/discovery-shared.ts
  git diff --check

• Evidence after fix:

  {
    "configuredApi": "openai-completions",
    "configuredBaseUrl": "http://192.168.20.14:11434/v1",
    "discoveredBuilderApi": "ollama",
    "discoveredBuilderBaseUrl": "http://192.168.20.14:11434",
    "resolvedProviderApi": "openai-completions",
    "resolvedBaseUrl": "http://192.168.20.14:11434/v1",
    "resolvedModelIds": [
      "qwen3-coder:cloud"
    ]
  }

  node scripts/run-vitest.mjs extensions/ollama/provider-discovery.test.ts
  Test Files  1 passed (1)
  Tests  18 passed (18)
  
  node scripts/run-vitest.mjs extensions/ollama/index.test.ts
  Test Files  1 passed (1)
  Tests  52 passed (52)
  
  pnpm check:test-types
  completed successfully
  
  node scripts/run-oxlint.mjs --deny=warn --ignore-path=.oxlintignore extensions/ollama/index.test.ts extensions/ollama/index.ts extensions/ollama/provider-discovery.test.ts extensions/ollama/src/discovery-shared.ts
  completed successfully
  
  git diff --check
  completed successfully
  

• Observed result after fix: The resolved Ollama provider keeps api: "openai-completions" and baseUrl: "http://192.168.20.14:11434/v1", while native discovery still probes the normalized Ollama API base at http://192.168.20.14:11434.
• What was not tested: Full gateway startup and a live external Ollama daemon were not used; the issue is a provider-local discovery merge regression and was verified through the real OpenClaw plugin discovery function plus focused Ollama plugin tests.

Regression Test Plan

• Coverage level: Provider plugin regression tests
• Target test file: extensions/ollama/provider-discovery.test.ts, extensions/ollama/index.test.ts
• Scenario locked in: Explicit models.providers.ollama.api: "openai-completions" and baseUrl: ".../v1" are preserved through catalog discovery and prepareDynamicModel/resolveDynamicModel even when buildOllamaProvider() returns native api: "ollama" and a native base URL.
• Why this is the smallest reliable guardrail: It covers the two Ollama-owned merge boundaries that previously replaced the explicit API/base URL without touching core routing or other providers.

Root Cause

• Root cause: buildOllamaProvider() correctly normalizes the configured base URL for native Ollama discovery, but the catalog and dynamic-model merge paths reused that native discovery provider as the final runtime provider/model config, replacing the user's OpenAI-compatible API/base URL pair.
• Missing detection / guardrail: Existing tests covered base URL normalization and dynamic model discovery separately, but did not assert that a user-configured OpenAI-compatible API and /v1 base URL survived discovery-backed paths.