Distillation

Date: Sep 17, 2026

• 🔴 (09:52) User requires invalid embedding-pool environment values to resolve to undefined, never NaN, and configuration parsing must not spin up workers.
• 🔴 (09:52) User requires the embedding pool’s primary worker to be always present; its own OOM backoff, rather than pool sizing, governs its recovery.
• 🔴 (09:52) User stated the native ONNX OOM can SIGKILL the entire gateway, so the ×0.7 backoff never fires; the only defense is preventing over-allocation in the first place.
• 🟡 (09:52) packages/core/test/embedding-pool-memory.test.ts documents Onur’s reported regression: the pool grew to N workers while each independently sized its token cap from roughly half of total free memory, causing aggregate host-memory exhaustion.
• 🟡 (09:52) The embedding-pool OOM fix has two guards: (1) each worker sizes its cap from free / ceiling; (2) every request re-clamps the cap against current free memory rather than retaining only its construction-time cap.
• 🟡 (09:52) packages/core/test/embedding-pool-memory.test.ts uses CapturingWorker extends EventEmitter to snapshot exact EmbedMsg payloads { type, id, maxTokens }; completeNext() emits one result vector new Float32Array([1, 0, 0]), and settle<T>() immediately installs fulfillment/rejection handlers so pending embeds cannot become unhandled rejections during teardown.
• 🟡 (09:52) Embedding-pool memory tests force the local provider by deleting VOYAGE_API_KEY and OPENAI_API_KEY, neutralize real CI cgroup limits with _setConstrainedMemoryForTest(0), and restore worker factory, pool-size seam, free-memory seams, constrained-memory seam, provider probe state, saved provider, and environment variables after every test.
• 🟡 (09:52) Pool-share regression test persists cap 8192 with freeMemBytes=0, sets _setEmbedPoolSizeForTest(2) and _setContainerFreeForTest(6 * GB), and requires the posted cap to equal memoryModelEmbedCap((6 * GB) / 2) and be strictly below memoryModelEmbedCap(6 * GB).
• 🟡 (09:52) Unknown-headroom tests require MIN_EMBED_TOKENS when _setAvailableMemoryForTest(Number.NaN) under both conditions: (1) constrained memory 12 * GB with container free 64 * GB; (2) no hard constraint (0) with container free 64 * GB.
• 🔴 (09:52) User requires unknown or failed memory-headroom accounting to fall back to the token floor rather than trusting apparently large container-free memory.
• 🔴 (09:52) User stated the current-free-memory re-clamp test uses one worker and the pool never grew because its ceiling is 1.
• 🟡 (09:52) Current-memory regression test persists cap 8192, sets pool ceiling 1, sends a first request at 6 * GB, completes it, drops free memory to 2 * GB, and sends a second request through the same worker; required caps are respectively memoryModelEmbedCap(6 * GB) and memoryModelEmbedCap(2 * GB), with the second strictly smaller.
• 🟡 (09:52) OOM-respawn regression test persists cap 8192, uses pool ceiling 1, begins at 16 * GB, drops free memory to 2 * GB, then emits EMBED_OOM_EXIT_CODE; the fresh worker must receive the same request ID with maxTokens === memoryModelEmbedCap(2 * GB), strictly below backoffEmbedCap(first.maxTokens).
• 🔴 (09:52) User requires OOM-respawn resubmission to clamp directly against current free memory, not merely apply the ×0.7 backoff, because this payload is posted directly without a later effectiveMaxTokens pass.
• 🟡 (09:52) packages/core/src/embedding.ts publicly re-exports pool test seams _configuredBackfillCpuDuty, _configuredEmbedPoolSize, _setEmbedPoolSizeForTest, _setEmbeddingWorkerWatchdogsForTest, and _setPoolFreememForTest; memory/provider seams include _setAvailableMemoryForTest, _setConstrainedMemoryForTest, _setContainerFreeForTest, and _setTestWorkerFactory.
• 🟡 (09:53) jj file annotate -L ... failed because jj file annotate does not accept -L; the CLI advised passing it as a value via -- -L and showed usage jj file annotate [OPTIONS] <PATH>.
• 🟡 (09:53) packages/core/test/setup.ts creates one isolated temporary database for the test run using mkdtempSync(join(tmpdir(), "lore-test-")) and sets LORE_DB_PATH to <tmp>/test.db, preventing fixtures from reaching ~/.local/share/lore/lore.db; afterAll calls close() and recursively removes the temporary directory.
• 🟡 (09:53) packages/core/test/setup.ts replaces globalThis.fetch with a baseline interceptor only for https://models.dev/api.json, returning canned HTTP 200 JSON while delegating all other requests to the real fetch. This prevents gateway pipeline pre-warms from depending on models.dev availability or polluting cross-file fetch-mock assertions.
• 🟡 (09:53) Canned models.dev entries are: claude-opus-4-6 with cost { input: 5, output: 25, cache_read: 0.5 } and limits { context: 1_000_000, output: 128_000 }; claude-sonnet-4-20250514 with { input: 3, output: 15, cache_read: 0.3 } and { context: 1_000_000, output: 64_000 }; claude-haiku-4-5 with { input: 1, output: 5, cache_read: 0.1 } and { context: 200_000, output: 64_000 }; gpt-5.4-mini with { input: 0.75, output: 4.5, cache_read: 0.19 } and { context: 400_000, output: 100_000 }.
• 🔴 (09:53) User requires stderr-silencing state to never cross a test boundary.
• 🟡 (09:53) Global afterEach in packages/core/test/setup.ts calls silenceStderr(false) because force-active Pi/OpenCode plugin tests using LORE_*_FORCE_ACTIVE=1 can otherwise leave shared-fork stderr silenced, hiding unrelated logs or vacuously satisfying “no output” assertions.
• 🔴 (09:53) User requires the per-connection project path→ID memo to be invalidated after every test so a settled or alias mapping cached by one test is never served to another test reusing the same path.
• 🟡 (09:53) Global afterEach in packages/core/test/setup.ts calls invalidateProjectIdCache() because the entire run shares one temporary database and one db() instance.
• 🟡 (09:53) packages/core/test/semantic-token-cache.test.ts verifies pending derived-cache publication is rejected after clearProject(scope.projectPath) and after session recreation, both through local deleteSession(...); parent() and through an external DatabaseSync transaction that deletes and reinserts the same session_id.
• 🟡 (09:53) Semantic-token-cache contention test sets caller PRAGMA busy_timeout = 1375, holds BEGIN IMMEDIATE on another DatabaseSync, requires pending.persist() to return in under 500 ms, restores timeout 1375, increments pending.stats.unavailable to 1, and leaves the prior payload unchanged; cleanup rolls back the external transaction and resets timeout to 5000.
• 🟡 (09:53) Semantic-token-cache transaction test uses SAVEPOINT test_cache_rollback, fills the cache, executes ROLLBACK TO test_cache_rollback; RELEASE test_cache_rollback, and requires the next fill to report stats.misses === 1.
• 🟡 (09:53) Semantic-token-cache merge test fills a source project, calls mergeProjectInternal(source, target), and verifies no cache row remains under the source project_id, preventing a merged project’s derived cache from being reused under another owner.