Skip to content

yg/phase a6 policy query tests#91

Open
couragehong wants to merge 2 commits intofeat/go-migrationfrom
yg/phase-a6-policy-query-tests
Open

yg/phase a6 policy query tests#91
couragehong wants to merge 2 commits intofeat/go-migrationfrom
yg/phase-a6-policy-query-tests

Conversation

@couragehong
Copy link
Copy Markdown
Contributor

@couragehong couragehong commented Apr 28, 2026 •
edited
Loading

Warning: This PR is created by Claude.

Summary

  • What: internal/policy/query_test.go + internal/domain/query_test.go 신규. 테스트 only, production 코드 변경 없음.
  • Why: policy/query.go (Python query_processor.py bit-identical 포팅) 와 domain/query.go 의 helpers (IsReliable/IsPhase/ExtractPayloadText) 에 테스트 0건. silent drift 시 recall accuracy 가 무성으로 떨어짐.
  • Scope: agents/tests/test_retriever.py::TestQueryProcessor 의 11개 중 10개 포팅 + Python 이 안 가드한 invariant 추가. 나머지 클래스는 Go 등가 코드 없음/D21/D28 로 모두 N/A.

Python 대비 매핑

Python 클래스 처리 사유
TestQueryProcessor (11) 10 포팅 test_format_for_search 만 N/A — Go 에 format_for_search 미존재
TestQueryProcessorMultilingual (8) N/A D21 — agent pre-translates, Go ParsedQuery 에 Language 필드 없음
TestSearcher / TestExpandPhaseChains N/A searcher Go 등가물 = service/recall.go (Phase 5, TODO 다수)
TestSynthesizer* (3 클래스) N/A D28 — agent-delegated, Go 에 synthesis 자체 없음
TestDisplayTextLocalization / FormatAnswer N/A display/synthesis 영역, recall 외부

강화 포인트 (Python 미가드)

  • 모든 QueryIntent 8개 / TimeScope 5개 망라 (Python 은 4 of 8, 2 of 5)
  • 규칙 iteration 순서 precedence (intent + time)
  • Cap 정확값 (len == 5/10/15, ≤ 아님)
  • AND assertion (Python 의 OR 보다 strict)
  • GENERAL intent fall-through contract
  • Stop word + length filter + dedup count exact-1
  • 1999 negative boundary (20\d{2} 패턴 강제)

Validation

  • go test -count=1 -race ./internal/policy/ ./internal/domain/ → ok (~1.4s, 27 함수 / 68 subtest)
  • gofmt -l / go vet → clean
  • 사전 리뷰 적용: 3 서브에이전트 (adversarial / Python parity / Go style) 가 잡은 HIGH 발견 모두 반영 (OR→AND, exact-cap, 규칙 precedence, slices.Contains, package doc 충돌 회피, table-driven IsPhase)

Cross-Agent Invariants

테스트 파일 2개만 추가. scripts/bootstrap-mcp.sh, agent 스크립트, Codex/Claude/Gemini/OpenAI 지시서, SKILL.md, commands/rune/*.toml, AGENT_INTEGRATION.md 모두 미수정 → 모든 invariant trivially 만족.

Notes for Reviewers

  • Risk: TestSearchHit_IsPhase/group_id_pointer_to_empty_string 가 현재 동작 (pointer 존재만 보고 IsPhase=true) 을 lock-in. 팀이 빈 문자열을 "phase 아님" 으로 결정하면 production IsPhase 와 같이 수정 필요. 코드에 TODO(yg) 표시.
  • BC: 없음 (test-only).
  • Follow-up: policy/{rerank,novelty,record_builder,payload_text,pii}.go, domain/schema.go (Python golden 비교 추천), lifecycle/shutdown.go::ZeroizeDEK (dead-store 최적화 가드 필요).

🤖 Generated with Claude Code

couragehong and others added 2 commits April 28, 2026 10:34
@couragehong @claude
test(policy,domain): unit tests for query.Parse + SearchHit/ExtractPa…
9f07d46 
…yloadText

Port the Python TestQueryProcessor suite (agents/tests/test_retriever.py)
to Go and extend coverage to every QueryIntent / TimeScope plus invariants
the Python suite did not gate explicitly.

internal/policy/query_test.go (black-box, package policy_test):
  - All 7 explicit intents + GENERAL fallback (was 5 of 8 in Python).
  - All 4 explicit time scopes + ALL_TIME default (was 2 of 5 in Python).
  - Entity extraction: quoted strings + capitalized + tech regex.
  - Keyword extraction: stop-word filter, length filter, dedup.
  - Query expansion: original included, intent variants, 5-cap.
  - Cleaning: lowercase, whitespace collapse, trim, trailing punctuation
    (question mark preserved; period/bang/comma/semicolon/colon stripped
    including consecutive runs).
  - Field caps: entities <= 10, keywords <= 15, expansions <= 5.

internal/domain/query_test.go:
  - SearchHit.IsReliable: supported/partially_supported true; everything
    else false (incl. empty string and unrelated values).
  - SearchHit.IsPhase: pointer presence drives the predicate (documented
    via a subtest, since pointer-to-empty-string also returns true).
  - ExtractPayloadText: standard path, missing payload, payload not a map,
    text not a string, text missing, nil metadata, payload nil.

Multilingual cases skipped: Go ParsedQuery has no Language field
(D21 - agent pre-translates), so the regex/LLM split that exists in
the Python QueryProcessor does not exist here.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@couragehong @claude
test(policy,domain): tighten assertions per pre-PR adversarial review
1efe251 
Address findings from a 3-agent pre-PR review (adversarial / Python parity /
Go style). Same scope as the previous commit; assertions made stricter so
plausible regressions are caught.

policy:
  - Replace OR with AND in TestParse_KeywordsRetainsContentTerms and
    TestParse_EntitiesCapitalizedAndTechPatterns. Half-broken extraction
    no longer slips through.
  - Cap tests assert exact equality (len == 5/10/15) on overflowing input,
    not just upper bound. A cap regression to a smaller value now fails.
  - Add TestParse_IntentRulePrecedence and TestParse_TimeRulePrecedence to
    gate the rule iteration order (silent reordering would be undetected
    by the per-intent / per-scope tests alone).
  - Add TestParse_EntityOrCleanedSurfacesToken — preserves the disjunction
    Python asserts in test_retriever.py:25 (entity verbatim OR token in
    cleaned), which the original Go split lost.
  - Add TestParse_ExpansionGeneralIntentNoPrefixes — gates the
    fall-through contract for intents not in generateExpansions's switch.
  - Tighten TestParse_ExpansionContainsCleanedAndIntentVariants: must use
    a query that actually matches DecisionRationale, and the assertion
    requires a prefix-form intent variant in the output.
  - Rename TestParse_KeywordsDeduplicated to
    TestParse_KeywordsDedupAfterLowercase — clarifies that the dedup is
    of already-lowercase strings (cleanQuery happens upstream).
  - Extend TestParse_TimeScope: this_week / this_quarter / past 3 months /
    past year / 1999 negative boundary (must NOT match 20\d{2}).
  - Extend TestParse_Cleaned: tab and newline whitespace collapse cases.

domain:
  - TestSearchHit_IsReliable: comment cites the 4 canonical Certainty
    values from Python schema and justifies the 6 chosen test cases.
  - TestSearchHit_IsPhase: convert to table-driven for consistency, mark
    pointer-to-empty-string with a TODO referencing the team decision.
  - TestExtractPayloadText: doc clarifies Go is intentionally narrower
    than Python (D32 — drops the 3 fallback paths), not a literal port.

Style:
  - Replace hand-rolled `contains` helper with `slices.Contains` (Go 1.25).
  - Move file-level header comments below `package` lines to avoid
    colliding with package doc comments.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@couragehong couragehong changed the title Yg/phase a6 policy query tests yg/phase a6 policy query tests Apr 28, 2026
@couragehong couragehong requested a review from esifea April 28, 2026 22:20
@couragehong couragehong mentioned this pull request Apr 29, 2026
Open
5 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@esifea esifea esifea approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@couragehong @esifea