Distillation

Date: Sep 16, 2026

• 🔴 (13:17) User specified request-local recall-diagnostic privacy with exact wording: “Fingerprints and recall content never leave this closure.” createRecallDiagnostics(enabled = true) keeps SHA-256 fingerprints only in closure-local Sets for inputs, results, pairs, and delivered coverage (packages/gateway/src/recall-diagnostics.ts, lines 7-20).
• 🟡 (13:17) createRecallDiagnostics() is bounded by MAX_RECALL_DIAGNOSTIC_ROUNDS = 64; record() stops when disabled, finished, or at the limit and tracks rounds, detailCalls, emptyBodies, resultBytes, repeated inputs/results/pairs, and novel coverage (packages/gateway/src/recall-diagnostics.ts, lines 4-60).
• 🟡 (13:17) Recall input fingerprints conservatively include scope ?? "all", id || ids || null, and the trimmed query only when no ID lookup is present. Coverage fingerprints combine identity, revision, kind ?? "detail", offset, and length with \u0000; zero-length coverage does not count as progress (packages/gateway/src/recall-diagnostics.ts, lines 34-58).
• 🟡 (13:17) Recall diagnostics emit recall-round through log.info() with round, kind (detail or search), repetition flags, coverageProgress, and per-result byte count. finish("completed" | "failed" | "aborted") is idempotent, emits recall-chain with aggregate counts and elapsed time clamped to 0..300_000 ms, then clears all four sets (packages/gateway/src/recall-diagnostics.ts, lines 59-75).
• 🟡 (13:17) Provider-failure work comprised 7 sequential changes: 1. d385b39653b0085833e281b6347b15071cee6801 / prpmozzzqvkt — fix(responses): sanitize provider failures; 2. 29dee093d7ceed38b0ed823c9d5b26f8c1ba3d7c / qnlnmzlxyynv — fix(responses): bind sanitized failures to trusted state; 3. b1bc7cd46c4eb09f23a7a5c6db257c380401ca02 / xqmskzvmnqkr — fix(responses): complete provider failure sanitization; 4. 01a6e06b08ba1b54f42c5b4c757ec449eb8e8490 / skwrwnvkvulp — fix(responses): make provider failure settlement atomic; 5. a1020a45be1382893a183e2ff39506e78a08fb76 / plzxzousrzuv — fix(responses): preserve trusted failure metadata; 6. 24880457763b6f10831856d01c41bd72d62ddc28 / wswlsmrnuzlq — fix(responses): reject non-monotonic provider failures; 7. 0dd569b520bad37adbaea230f32100c7f83f463c / lpmvqrurxwlm — fix(responses): unify provider failure handling.
• 🟡 (13:18) Final provider-failure commit 0dd569b520bad37adbaea230f32100c7f83f463c carries bookmarks fix/responses-provider-failures and fix/responses-provider-failures@git; it changed packages/gateway/src/pipeline.ts, packages/gateway/src/stream/openai-responses.ts, packages/gateway/test/openai-responses-recall-aware-stream.test.ts, and packages/gateway/test/openai-responses-stream.test.ts with 364 insertions and 27 deletions.
• 🟡 (13:19) Provider-failure handling was unified around exported validateResponsesProviderFailureTerminal() in packages/gateway/src/stream/openai-responses.ts. It requires parsed.type === event and trusted state.created; validates the sanitized response.failed projection using validatePublicResponsesEvent() or validateCodexResponsesEvent(); and additionally calls validateFailureTerminal(parsed) for events other than diagnostic error and response.error.
• 🟡 (13:19) Both primary and continuation provider-failure paths in packages/gateway/src/pipeline.ts now call validateResponsesProviderFailureTerminal() before lifecycle validation and settlement, using validation mode opts.validation ?? "public" and sparse bound Math.min(maxSSEFrames, DEFAULT_MAX_SSE_FRAMES).
• 🟡 (13:19) Every accepted OpenAI Responses provider failure now emits the constant RESPONSES_PROVIDER_FAILURE_LOG = "openai-responses provider failure" via always-visible log.error(), including accumulator, main pipeline, and continuation paths.
• 🟡 (13:19) New projectResponsesInProgressEvent(parsed, state) constructs externally visible response.in_progress frames from trusted response.created state: trusted id, model, optional created_at, fixed status: "in_progress", fixed output: [], object normalized to "response", and optional source sequence_number. Both plain accumulation callbacks and recall-aware forwarding use this projection instead of forwarding provider lifecycle metadata verbatim.
• 🟡 (13:19) OpenAI Responses model-consistency validation now exempts response.in_progress source metadata because that frame is projected from trusted created state before exposure; other events still reject a model differing from state.model.
• 🟡 (13:19) validateFailureTerminal() now also validates parsed.response.usage through validateResponsesUsage(parsed.response.usage, "malformed Responses usage"), preserving validated failure usage while rejecting malformed usage.
• 🟡 (13:21) Tests explicitly cover trusted-state and failed-output behavior: packages/gateway/test/openai-responses-stream.test.ts includes “uses created metadata and failure usage in the plain streaming route” (line 618) and “binds a sanitized failure to created state and clears failed terminal output” (line 4722); packages/gateway/test/openai-responses-recall-aware-stream.test.ts includes “uses created metadata and clears failed output from completion state” (line 646) and “keeps created metadata authoritative through a hidden recall and provider failure” (line 829).
• 🟡 (13:22) accountUnsuccessfulResponse() in packages/gateway/src/pipeline.ts records validated usage without publishing successful-turn state. It defaults missing usage to ZERO_USAGE, only accounts when at least one token value is numeric and greater than 0, invokes optional markDirty(), calls accountConversationUsage(), adds GenAI usage attributes when a span exists, then always sets span status { code: 2, message: "upstream response did not complete" } and calls endSpan() (lines 13479-13512).
• 🟡 (13:22) conversationTTLForAccounting() first reads the live session’s resolvedConversationTTL, then persisted loadSessionTracking(sessionID)?.resolvedConversationTTL, accepting only "5m" or "1h" and otherwise returning undefined (packages/gateway/src/pipeline.ts, lines 13514-13523).
• 🟡 (13:23) vitest.config.ts keeps resolve.alias at the top level because placing aliases under test.resolve.alias silently resolved stale dist builds and masked real failures. Aliases are @loreai/core → packages/core/src and @loreai/gateway → packages/gateway/src.
• 🟡 (13:23) Main Vitest configuration includes packages/core/test/**/*.test.ts, packages/core/eval/**/*.test.ts, packages/gateway/test/**/*.test.ts, packages/opencode/test/**/*.test.ts, and packages/pi/test/**/*.test.ts; uses setup file ./packages/core/test/setup.ts, Node environment, forks pool, testTimeout: 300_000, and hookTimeout: 300_000.
• 🟡 (13:23) Vitest test environment sets NODE_ENV: "test", SENTRY_ENABLED: "0", and LORE_DEBUG: "0". V8 coverage reporters are text, json, html, and lcov; exclusions include tests, setup/helpers, scripts, distributions, website, Pi, eval files, config files, and scripts/**.
• 🟡 (13:24) packages/core/test/setup.ts creates one run-local temporary directory using mkdtempSync(join(tmpdir(), "lore-test-")), sets LORE_DB_PATH to <tmp>/test.db, and after all tests calls close() then recursively removes the directory.
• 🟡 (13:24) Test setup intercepts only https://models.dev/api.json, returning canned HTTP 200 JSON, while delegating all other requests to captured realFetch. This prevents gateway pre-warming from making live calls, avoids third-party 500/timeout flakes, and prevents cross-file pollution of worker-model.test.ts fetch mocks.
• 🟡 (13:24) Canned models.dev data defines: Anthropic claude-opus-4-6 with costs { 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 }; OpenAI gpt-5.4-mini with { input: 0.75, output: 4.5, cache_read: 0.19 } and { context: 400_000, output: 100_000 }.
• 🔴 (13:24) User specified that plugin-induced stderr silencing “never crosses a test boundary.” afterEach() calls silenceStderr(false) so force-active Pi/OpenCode plugin tests cannot hide logs or vacuously satisfy unrelated “no output” assertions in the same fork (packages/core/test/setup.ts, lines 83-90).
• 🔴 (13:24) User specified that a settled or alias project path→ID mapping cached by one test “must never be served to another test that reuses the same path.” afterEach() calls invalidateProjectIdCache() because the whole run shares one temporary DB and one db() connection (packages/core/test/setup.ts, lines 90-95).
• 🔴 (13:26) User specified the mutation-testing review rule with exact wording: “ALWAYS hand-verify a survivor by applying the EXACT replacement from the report.” Stryker may split compound A && B expressions, so a reported :col → "true" mutation can target only a sub-operand rather than the entire condition; supporting guidance is in quality/MUTATION_TESTING.md.
• 🔴 (13:26) User specified that mutation testing “never fails CI — we record the score and ratchet over time.” Mutation scores are advisory while establishing the issue #832 baseline; thresholds are { high: 80, low: 60, break: null }, and runs are nightly/on-demand rather than per-PR.
• 🟡 (13:26) Stryker mutation testing targets high-risk stateful areas where review caught bugs missed by tests: sync prune-floor wedge and tier residue (#828) plus decay/evict edge cases (#816). Current mutate allowlist, in order, is 1. packages/core/src/sync-data.ts, 2. packages/gateway/src/sync.ts, 3. packages/gateway/src/translate/bedrock.ts; ltm.ts and gradient.ts are deferred because their coverage spans many test files.
• 🟡 (13:26) stryker.config.mjs explicitly loads @stryker-mutator/vitest-runner because plugin auto-discovery fails under pnpm’s non-hoisted symlink layout; it uses testRunner: "vitest", vitest.configFile: "vitest.mutation.config.ts", coverageAnalysis: "perTest", concurrency: 2, timeoutMS: 60_000, and timeoutFactor: 2.5.
• 🟡 (13:26) Mutation reporters are html, clear-text, progress, and json; HTML output is reports/mutation/index.html, with JSON at reports/mutation/mutation.json. Commands documented are pnpm mutation for the whole allowlist and pnpm mutation -- --mutate "packages/core/src/sync-data.ts" for one module.
• 🟡 (13:26) Stryker sandbox exclusions are dist, **/dist/**, dist-bin, dist-tarballs, dist-vendor, reports, coverage, .stryker-tmp, and **/*.tsbuildinfo.