feat(cli-pr-monitor): cli-finding-classifier 統合 + Finding C strict (ADR-038 Phase 5)

[aloekun]

Summary

ADR-038 (試験運用) Phase 5 として、cli-pr-monitor poll stage に cli-finding-classifier.exe を subprocess 経由で統合。同時に PR #119 の CR review round 3 で再指摘された Finding C (normalized_issue 契約検証) を strict 化。

• A. Phase 5 統合: 新 module classifier_runner.rs + ClassifierConfig (default OFF / 試験運用 flag) + state.classified_findings field + enrich_with_classifier step
• B. Finding C strict: from_llm_output で trim 後に「改行」「>80 chars」を検出し fallback に倒す + NORMALIZED_ISSUE_MAX_CHARS=80 const 化

設計のポイント

• block しない設計: classifier 失敗 (exe 不在 / spawn / stdin / timeout / parse) はすべて空 Vec を返し、cli-pr-monitor の poll は影響を受けない。ADR-038 §「失敗時の振る舞い」に整合
• 疎結合 (subprocess 経由): cli-pr-monitor の依存ツリーに ureq / Ollama を持ち込まず、cli-finding-classifier.exe を spawn する形で統合。schema (ClassifiedFinding) のみ複製し、ADR-038 で schema 変更が起きた際は両方を同期する責務をコメントで明示
• 後方互換: state.classified_findings は #[serde(default, skip_serializing_if = "Vec::is_empty")] で既存 state.json と consumer (takt facets / Claude) を破壊しない
• デフォルト OFF: ClassifierConfig.enabled = false。dogfood 段階で個別に opt-in する設計
• iteration 間の保持: build_state_for_iteration で existing.classified_findings を引き継ぎ (takt auto-fix iter で追加)

検証結果

• cli-finding-classifier: lib 15 pass (B 関連 +3 件) + bin 6 pass
• lib-ollama-client: 8 pass
• cli-pr-monitor: 148 pass (classifier_runner 4 + ClassifierConfig 2 含む)
• workspace 全体 clippy --all-targets -- -D warnings clean
• takt pre-push-review: APPROVE (2 iter, 14m 26s, convergence_verdict: fully_resolved)

scope 外 (本 PR では未対応)

• prompt injection サニタイズ: 本 PR は classifier を invoke して state に保存するだけで、auto_fix execution の自動実行経路を作っていない。Claude / takt が後段で human_review に倒れた finding を判断する設計のため、現時点で実害なし。auto_fix を機械的に実行する経路を追加するタイミングで sanitization を導入予定 (PR feat(cli-finding-classifier): Ollama でローカル LLM 分類 CLI 追加 (ADR-038) #119 の WARN ベース)
• ADR-038 line 61 textual fix: confidence=0.0 → action_confidence=0.0 の名称統一は別 PR の cleanup で対応 (機能影響なし)

Test plan

• cargo test --workspace --lib --bins 全 pass
• cargo clippy -p cli-pr-monitor -p cli-finding-classifier -p lib-ollama-client --all-targets -- -D warnings clean
• pnpm build:all 成功 (.claude/cli-finding-classifier.exe 配置済、cli-pr-monitor.exe も再ビルド)
• takt pre-push-review APPROVE
• (post-merge dogfood) [classifier] enabled = true で 1 PR の review サイクルを実走、token 削減効果を実測 — 別タスクで実施

関連

• ADR-038 (試験運用): Phase 5 の land で「試験運用 → 本採用への昇格条件」 2 番目を満たす
• docs/local-llm-offload-analysis.md §8.A (Phase 5 candidate)、§8.B (Finding C strict)
• PR feat(cli-finding-classifier): Ollama でローカル LLM 分類 CLI 追加 (ADR-038) #119: 提案 2 初版 land、本 PR の前提

Summary by CodeRabbit

• 新機能

  • 検出結果の自動分類器を統合しました。ポーリング処理内で分類を実行し、分類結果を保持・更新します。

• 改善

  • 正規化された問題説明の検証を強化し、最大長と単一行制約を追加しました。

• 設定

  • 分類器の設定（有効化、モデル、エンドポイント、タイムアウト）を追加しました。既定値あり。

• ドキュメント

  • 実装進捗ログと提案2の詳細、進捗・検証結果を追記しました。

[coderabbitai]

Warning

Rate limit exceeded

@aloekun has exceeded the limit for the number of commits that can be reviewed per hour. Please wait 12 minutes and 53 seconds before requesting another review.

To continue reviewing without waiting, purchase usage credits in the billing tab.

⌛ How to resolve this issue?

After the wait time has elapsed, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

We recommend that you space out your commits to avoid hitting the rate limit.

🚦 How do rate limits work?

CodeRabbit enforces hourly rate limits for each developer per organization.

Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout.

Please see our FAQ for further information.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: ee304013-b70c-4c75-b0d1-c69fc94a06b8

📥 Commits

Reviewing files that changed from the base of the PR and between fc07755 and 67e64bd.

📒 Files selected for processing (2)

• src/cli-pr-monitor/src/classifier_runner.rs
• src/cli-pr-monitor/src/stages/poll.rs

📝 Walkthrough

Walkthrough

LLM分類器をpr-monitorに統合。ClassifierConfig追加、CLI分類器ランナー新規実装、ポーリングループへのエンリッチ処理、PrMonitorStateの永続化フィールド追加、分類器側のnormalized_issue検証強化、および関連ドキュメント更新を行う。

Changes

分類器統合と検証強化

Layer / File(s)	Summary
設定スキーマ src/cli-pr-monitor/src/config.rs	ClassifierConfigを追加し、Configにclassifier: ClassifierConfigフィールドを導入。デフォルト値（enabled=false, model="mistral:7b", endpoint="http://localhost:11434", timeout_secs=30）を実装。テスト追加。
状態スキーマ src/cli-pr-monitor/src/state.rs	PrMonitorStateにclassified_findings: Vec<ClassifiedFinding>フィールドを追加。serde属性と初期化を追加し、テスト初期化を更新。
ランナー実装（コア） src/cli-pr-monitor/src/classifier_runner.rs	新規モジュール。ClassifiedFinding型、classifier_exe_path()、classify_findings()を実装。CLI実行ファイルの解決、JSON入出力、タイムアウト、子プロセス管理、失敗時は空Vecを返す（ADR-038に準拠）。ユニットテストを追加。
ポーリングループ統合 src/cli-pr-monitor/src/stages/poll.rs	PollContextにclassifier_configを注入。各イテレーション後にenrich_with_classifier()を呼び出し、既存のclassified_findingsを保持・更新して状態を永続化。テストのコンテキスト初期化を更新。
既存モジュール参照 src/cli-pr-monitor/src/main.rs	mod classifier_runner;を追加して新モジュールをコンパイル対象にする。
分類器ライブラリの検証強化 src/cli-finding-classifier/src/lib.rs	normalized_issueの長さ上限を80文字に設定。マルチラインや超過長の入力に対してフォールバック理由を設定し、検証ロジックと関連テストを追加。
ドキュメント更新 docs/local-llm-offload-analysis.md	ステータス表現を更新し「実装進捗ログ」セクションを追加、ADR-038参照や検証結果・進捗記録を記載。旧Next Stepsを置換。

Sequence Diagram

sequenceDiagram
    participant Poll as ポーリングループ
    participant Runner as ランナー
    participant Exe as CLI分類器

    Poll->>Runner: classify_findings(config, findings)
    Runner->>Runner: 入力検証・早期リターン
    alt 入力有効かつ実行ファイル解決成功
        Runner->>Exe: JSONをstdinで送信
        Exe->>Exe: 分類処理（外部CLI）
        Exe-->>Runner: JSON出力
        Runner->>Runner: 出力をパース・Vec<ClassifiedFinding>を生成
        Runner-->>Poll: 分類結果を返却
    else 失敗（例: missing exe, timeout, parse error）
        Runner-->>Poll: 空のVecを返却
    end
    Poll->>Poll: classified_findings を更新・永続化

Loading

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

• aloekun/claude-code-hook-test#38: モジュール分割を導入したPRと同じcli-pr-monitorファイル群を変更しており、今回のclassifierランナー追加がそのモジュール構造に依存している。
• aloekun/claude-code-hook-test#97: ConfigおよびPrMonitorStateへのフィールド追加やpollステージ改修と重複する変更セットを含むため関連。
• aloekun/claude-code-hook-test#114: pollループの制御や状態フィールドを変更するPRで、ポーリングフローへの機能追加と衝突／相互影響の可能性が高い。

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	タイトルは主な変更内容（cli-finding-classifier の統合と Finding C の厳密化）を明確に説明しており、ADR-038 Phase 5 というフェーズ参照も含まれています。変更セットの核心的な内容をよく要約しています。
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[aloekun]

@coderabbitai review

[coderabbitai]

✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[aloekun]

@coderabbitai review

[coderabbitai]

✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

Actionable comments posted: 2

🧹 Nitpick comments (2)

src/cli-pr-monitor/src/stages/poll.rs (1)

109-109: 💤 Low value

iteration ごとの classifier 呼び出しコスト。

run_one_iteration 毎に enrich_with_classifier が走ると、wakeup 経路で複数 iteration がある場合に Ollama を都度叩くことになります。現状は classified_findings 非空ガードで実質 1 回しか走らない実装ですが、上のコメントでガードを外すと毎 iteration で findings.len() 件の Ollama 呼び出しが発生します (timeout 30s × N findings)。

Phase 5 試験運用としては許容範囲ですが、本採用前に「findings の内容 hash 変化時のみ再分類」程度のキャッシュ戦略を検討する価値があります。本 PR で対応必須ではないので参考までに。

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/cli-pr-monitor/src/stages/poll.rs` at line 109, The current code calls
enrich_with_classifier(&mut state, ctx.classifier_config) every
run_one_iteration which can re-query Ollama per finding; implement a simple
cache keyed by a content hash to avoid reclassification unless a finding's
content changed: add a classification cache map on the state (e.g.
state.classification_cache: HashMap<ContentHash, Classification>), compute a
stable hash for each finding payload before calling enrich_with_classifier in
run_one_iteration, filter/find only changed findings (or skip calling
enrich_with_classifier entirely if none changed), update the cache with new
classifications after enrich_with_classifier returns, and use
state.classification_cache lookups to populate classified_findings so behavior
is unchanged while avoiding redundant Ollama calls.

src/cli-pr-monitor/src/classifier_runner.rs (1)

38-47: 💤 Low value

current_exe() 失敗時に CWD 配下を探索する挙動が意図通りか確認したいです。

unwrap_or_default() で PathBuf::new() (空パス) を返した後、parent() は None、unwrap_or(Path::new(".")) で . に倒れて ./cli-finding-classifier.exe を返します。exists() チェックがあるので致命的ではないですが、本来想定していない CWD 配下の同名 exe を誤って起動する余地が残ります (cli-pr-monitor は通常 .claude/ から起動される前提)。

Windows-only かつ current_exe() が失敗するのは極めて稀なため `` です。runner.rs::checker_exe_path と挙動を揃えるだけで済むので、必要なら近傍 PR でまとめて対応するのが妥当です。

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/cli-pr-monitor/src/classifier_runner.rs` around lines 38 - 47, The
current implementation uses unwrap_or_default() which makes current_exe()
failures silently fall back to CWD (./cli-finding-classifier.exe); change
classifier_exe_path() to only use the executable's parent if current_exe()
succeeds and otherwise use the same fallback logic as
runner.rs::checker_exe_path to avoid accidentally picking up a CWD
binary—specifically replace the unwrap_or_default()/parent()/unwrap_or(...)
chain with current_exe().ok().and_then(|p| p.parent().map(|d|
d.join("cli-finding-classifier.exe"))).unwrap_or_else(|| /* use the same
fallback PathBuf used in checker_exe_path */) so the behavior matches
checker_exe_path().

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/cli-pr-monitor/src/classifier_runner.rs`:
- Around line 141-207: The child-process handling currently polls in
wait_with_timeout without draining stdout/stderr, risking pipe-buffer deadlock;
update the implementation so stdout/stderr are consumed while the child runs
(either spawn threads to read_to_end for child.stdout and child.stderr and then
construct an Output when the process exits inside wait_with_timeout, or change
the child's stderr to inherit/null to eliminate the pipe risk); locate usages in
wait_with_timeout and feed_stdin/parse_classifier_output to ensure you read the
pipes you created with Stdio::piped() (or change Stdio::stderr to inherit/null
where the child is spawned) so the classifier never blocks writing large stderr
output.

In `@src/cli-pr-monitor/src/stages/poll.rs`:
- Around line 222-236: The early-return guard in enrich_with_classifier prevents
new findings from being classified because it returns when
state.classified_findings is non-empty; remove the
"!state.classified_findings.is_empty()" check so enrich_with_classifier only
skips when config.enabled is false or state.findings is empty, call
classify_findings(config, &state.findings) each iteration, assign the result to
state.classified_findings and persist via write_state as currently done; ensure
this change aligns with build_state_for_iteration and
update_state_from_check_result semantics (or alternatively add explicit logic to
clear state.classified_findings when findings change if you prefer preserving
prior classifications).

---

Nitpick comments:
In `@src/cli-pr-monitor/src/classifier_runner.rs`:
- Around line 38-47: The current implementation uses unwrap_or_default() which
makes current_exe() failures silently fall back to CWD
(./cli-finding-classifier.exe); change classifier_exe_path() to only use the
executable's parent if current_exe() succeeds and otherwise use the same
fallback logic as runner.rs::checker_exe_path to avoid accidentally picking up a
CWD binary—specifically replace the unwrap_or_default()/parent()/unwrap_or(...)
chain with current_exe().ok().and_then(|p| p.parent().map(|d|
d.join("cli-finding-classifier.exe"))).unwrap_or_else(|| /* use the same
fallback PathBuf used in checker_exe_path */) so the behavior matches
checker_exe_path().

In `@src/cli-pr-monitor/src/stages/poll.rs`:
- Line 109: The current code calls enrich_with_classifier(&mut state,
ctx.classifier_config) every run_one_iteration which can re-query Ollama per
finding; implement a simple cache keyed by a content hash to avoid
reclassification unless a finding's content changed: add a classification cache
map on the state (e.g. state.classification_cache: HashMap<ContentHash,
Classification>), compute a stable hash for each finding payload before calling
enrich_with_classifier in run_one_iteration, filter/find only changed findings
(or skip calling enrich_with_classifier entirely if none changed), update the
cache with new classifications after enrich_with_classifier returns, and use
state.classification_cache lookups to populate classified_findings so behavior
is unchanged while avoiding redundant Ollama calls.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 503ee52a-d74f-425b-8819-7d933f9a87d1

📥 Commits

Reviewing files that changed from the base of the PR and between 9f368a2 and fc07755.

📒 Files selected for processing (7)

• docs/local-llm-offload-analysis.md
• src/cli-finding-classifier/src/lib.rs
• src/cli-pr-monitor/src/classifier_runner.rs
• src/cli-pr-monitor/src/config.rs
• src/cli-pr-monitor/src/main.rs
• src/cli-pr-monitor/src/stages/poll.rs
• src/cli-pr-monitor/src/state.rs

[aloekun]

@coderabbitai review

[coderabbitai]

✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.