Resolves #14009
Related: #13999
Fixes the actual LM Studio warning path. TextEmbeddingService now preserves LMS loaded-model metadata from lms ps --json and, only when the active OpenAI-compatible endpoint is the orchestrator-owned LMS lane and the resident embedding model metadata identifies GGUF/Qwen3, appends <|im_end|> to the outbound embedding request strings. Stored Memory Core / KB text is not mutated, and generic OpenAI-compatible or Ollama-compatible endpoints do not receive the LMS/Qwen suffix.
Evidence: L2 unit/preflight evidence covers the request boundary; live local LMS metadata was also checked and shows format: gguf, architecture: qwen3, and embeddingModel: text-embedding-qwen3-embedding-8b on the configured LMS port. I did not add another live embedding request while LMS reported queued work.
Deltas from ticket
The first version of this PR was wrong: it documented or tested around the warning instead of satisfying LMS. This revision implements the ticket goal directly by making Neo send strings whose final token resolves to the LMS-required Qwen3 separator on the LMS metadata path. The earlier cloud DeploymentCookbook note is gone; cloud deployments do not use macOS lms.
Test Evidence
npm run test-unit -- test/playwright/unit/ai/services/memory-core/TextEmbeddingService.retry.spec.mjs test/playwright/unit/ai/scripts/runners/runSandman.spec.mjs — 85 passednpm run agent-preflight -- ai/services/graph/providerReadinessHelper.mjs ai/services/memory-core/TextEmbeddingService.mjs test/playwright/unit/ai/services/memory-core/TextEmbeddingService.retry.spec.mjs test/playwright/unit/ai/scripts/runners/runSandman.spec.mjsgit diff --check- Local LMS/config sanity: configured host
http://127.0.0.1:1234, orchestrator.lms.enabled=true, LMS port 1234; lms ps --json exposes format: gguf, architecture: qwen3.
Post-Merge Validation
Commits
64274a27d2 — fix(ai): append LMS embedding SEP from metadata (#14009)
Authored by Euclid (GPT-5 Codex, Codex Desktop). Session 9280140f-8b54-4462-9342-49cca7e226f4.
Resolves #14009
Related: #13999
Fixes the actual LM Studio warning path.
TextEmbeddingServicenow preserves LMS loaded-model metadata fromlms ps --jsonand, only when the active OpenAI-compatible endpoint is the orchestrator-owned LMS lane and the resident embedding model metadata identifies GGUF/Qwen3, appends<|im_end|>to the outbound embedding request strings. Stored Memory Core / KB text is not mutated, and generic OpenAI-compatible or Ollama-compatible endpoints do not receive the LMS/Qwen suffix.Evidence: L2 unit/preflight evidence covers the request boundary; live local LMS metadata was also checked and shows
format: gguf,architecture: qwen3, andembeddingModel: text-embedding-qwen3-embedding-8bon the configured LMS port. I did not add another live embedding request while LMS reported queued work.Deltas from ticket
The first version of this PR was wrong: it documented or tested around the warning instead of satisfying LMS. This revision implements the ticket goal directly by making Neo send strings whose final token resolves to the LMS-required Qwen3 separator on the LMS metadata path. The earlier cloud DeploymentCookbook note is gone; cloud deployments do not use macOS
lms.Test Evidence
npm run test-unit -- test/playwright/unit/ai/services/memory-core/TextEmbeddingService.retry.spec.mjs test/playwright/unit/ai/scripts/runners/runSandman.spec.mjs— 85 passednpm run agent-preflight -- ai/services/graph/providerReadinessHelper.mjs ai/services/memory-core/TextEmbeddingService.mjs test/playwright/unit/ai/services/memory-core/TextEmbeddingService.retry.spec.mjs test/playwright/unit/ai/scripts/runners/runSandman.spec.mjsgit diff --checkhttp://127.0.0.1:1234,orchestrator.lms.enabled=true, LMS port1234;lms ps --jsonexposesformat: gguf,architecture: qwen3.Post-Merge Validation
tokenizer.ggml.add_eos_token should be set to true.Commits
64274a27d2—fix(ai): append LMS embedding SEP from metadata (#14009)Authored by Euclid (GPT-5 Codex, Codex Desktop). Session 9280140f-8b54-4462-9342-49cca7e226f4.