gemma4:26b 가 "기기기기..." repetition loop 에 빠져 num_predict cap 도달 →
JSON truncate → unparseable. 두 가지 fix:
- Ollama body 에 repeat_penalty: 1.15 추가 (token repetition 억제)
- parseJsonLoose fail 시 throw 대신 {} 반환 → schema graceful coerce 가
placeholder title/summary 채움. raw_text 는 보존 → 사용자 데이터 무손실.
- coerceNullable 가 undefined 도 처리 (빈 객체 케이스).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
vision model 의 markdown fence / prose 섞인 응답에서 JSON 추출 fallback.
prompt 에 title 한국어 / kebab tags / JSON-only 출력 명시 강화.
- LocalOllamaProvider: parseJsonLoose 헬퍼 (첫 { ~ 마지막 } 추출)
- visionPrompt: 4 규칙 + markdown fence 금지 명시
- 단위 +2 (fence 추출 + prose 추출)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
err.message 안에 LAN endpoint URL (예: 192.168.x.x:11434) 이 포함될 수
있어 telemetry 파일에 PII 우회 노출. v0.2.3.1 in-app endpoint UI 가 LAN
사용을 흔하게 만들어 노출 경로 확대.
classifyFetchError 로 error class 분류 (network/timeout/dns/other) 후
reason: 'unreachable:{class}' 형태만 emit. host/IP 노출 0.
- abortController 가 method-local 에서 private instance field 로 이동
- public abort() 메서드 — 외부에서 in-flight generate 강제 중단
- ProviderHolder.replace() 시 호출되어 endpoint 변경 즉시 반영
- 단위 +2 cases (abort cancellation, model 파라미터)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- GenerateInput.vocab?: string[] (optional, 미전달 시 빈 배열 처리)
- LocalOllamaProvider.generate 가 input.vocab ?? [] 를 buildPrompt 4th 인자로
- 단위 +1 case (vocab → prompt body)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Flow 반전 (F7-D 채택):
- 기존: rule.iso ?? ai.dueDate (rule 우선)
- 신규: ai.dueDate ?? null (AI 우선)
규칙은 parseAllCandidates 로 모든 매치를 추출 → prompt 에 후보
힌트로 주입. AI 가 종합 판단. AI 실패 시 due_date null (별 fallback 없음).
해결되는 케이스: '내일 모레' → AI 가 ambiguous 인지 → null.
PROMPT_VERSION → 3. GenerateInput.dueDateCandidates 신규.
buildPrompt(rawText, todayKst, candidates) — 빈 배열일 때 hint 섹션 생략.
Tests:
- AiWorker.test.ts — 'rule priority' 테스트 → 'AI dueDate wins' flip
- AiWorker.test.ts — passes todayKst 테스트 확장 (dueDateCandidates 도 검증)
- AiWorker.test.ts — 신규 'passes parseAllCandidates result as dueDateCandidates'
- LocalOllamaProvider.test.ts / ollama-golden.test.ts — generate 호출에 dueDateCandidates: [] 추가
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
GenerateInput gains todayKst field. AiWorker computes KST-aligned
date once per job, runs parseDueDate on rawText, calls provider.generate
with todayKst, then merges: rule.iso wins if matched (deterministic),
else AI's due_date, else null. Logs dueDateSource (rule|ai|none) for
debugging. now() injection for testability.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Task 12 of the slice plan. Implements the slice's only provider:
- generate(): POST {endpoint}/api/generate with Korean-first
prompt; AbortController-driven 120s timeout; parses
body.response as JSON and runs it through parseAiResponse.
- healthCheck(): GET /api/tags; returns ok when configured model
is in the listing, otherwise reports the missing-model reason.
- Constructor takes opts.endpoint / opts.model so Task 30 main
entry can inject INKLING_OLLAMA_ENDPOINT for LAN dogfood;
defaults are http://localhost:11434 and gemma4:e4b.
Tests: 6 unit cases via undici MockAgent (parse, non-JSON,
timeout abort, healthCheck ok / missing / connection error).
Integration test gated by INKLING_INTEGRATION=1 hits real
Ollama for Korean / English-stack / mixed input cases with a
180s per-test budget — also reads INKLING_OLLAMA_ENDPOINT so
LAN dogfood can be exercised end-to-end.
Plan deviation: undici@8 types treat MockAgent's `.reply()`
callback as sync-return-only, so the 500ms-delayed reply used
in the timeout test is cast `as never` to bypass the overload
mismatch. Behavior is correct at runtime; the cast is local to
the test.
Verification: `npx vitest run tests/unit/LocalOllamaProvider.test.ts`
6 passed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
