feat: add compatible turn boundary markers

[fangxiu-wf]

Summary

• add a centralized fill-only TurnBoundaryProcessor before output dispatch
• cover canonical root turns that have a reliable gen_ai.turn.id, while preserving producer-owned markers
• keep Turn Boundary deduplication bounded and process-local, without participating in input checkpoints
• strengthen the JSONL validator so complete turns missing required boundaries are rejected

Compatibility guarantees

• existing Qoder, WorkBuddy, and Codex markers remain authoritative and unchanged
• Codex completed/interrupted and OpenClaw llm_output terminal semantics are preserved
• subagent records and supplemental telemetry without turn.id are skipped
• no records are added, removed, merged, or reordered
• event IDs, timestamps, trace context, token/message/tool data, finish reasons, OTLP span topology, and flush behavior are unchanged
• Turn Boundary enrichment does not access StateStore or input offsets/row IDs, and does not modify Agent inputs, deployment, or Flusher implementations
• enrichment failures are fail-open

Accepted first-stage boundaries

• events without a reliable gen_ai.turn.id remain unchanged
• turns without a reliable terminal signal do not receive a guessed gen_ai.turn.end
• process-local state resets when Pilot restarts, so a turn spanning a restart may receive another gen_ai.turn.start

Validation

• npm run typecheck
• npm run build
• focused Turn Boundary/InputManager/OpenClaw/validator regression: 83 passed
• StateStore regression: 17 passed
• OTLP trace flusher regression: 86 passed
• full local suite: 3263 passed, 50 skipped, 0 failed
• git diff --check

Fixes #261

[ralf0131]

Summary

Centralized fill-only TurnBoundaryProcessor that enriches entries with reliable gen_ai.turn.start/end markers before output dispatch. Well-designed with proper fail-open semantics, producer-marker preservation, bounded state tracking, and comprehensive test coverage across unit/integration/e2e layers.

Findings

• [Info] src/normalization/turn-boundary-processor.ts:11 — TERMINAL_FINISH_REASONS is duplicated in scripts/e2e/lib/e2e-scenarios.mjs:797. Consider extracting to a shared constant if the e2e validator needs to stay in sync as terminal reasons evolve.
• [Info] src/normalization/turn-boundary-processor.ts — persist() is called on every enrich() batch. For high-throughput inputs this could add I/O overhead. The current design is correct; just flagging for awareness if profiling ever surfaces this as a hotspot.

Suggestions

The implementation correctly handles the key edge cases:

• Subagent entries are skipped (no false boundaries)
• Producer-owned markers are never overwritten (inspect-before-fill pattern)
• StateStore restore validates version and TTL
• compact() caps at 4096 tracked turns with FIFO eviction

LGTM — clean, well-tested, backward-compatible change.

---

Automated review by github-manager-bot

[linrunqi08]

🔍 Code Review Summary

Severity	Count
Critical	0
High	0
Medium	0
Low	3

Lifecycle Verdict

Check	Result
资源释放	PASS
死锁/卡死风险	PASS
状态恢复正确性	PASS
PowerShell CLM 合规	N/A

无 FAIL 项。补标器无 timer/句柄；enrich() 同步且 fail-open；跟踪故意 process-local，不写 StateStore。

Merge Gate（合入门禁）

APPROVE-READY ✅（阻断级别：Critical / High / Medium；Low 不阻断；Lifecycle 任一 FAIL 也阻断）

本轮无 open 阻断项。

总体结论

中央 fill-only TurnBoundaryProcessor 接在 InputManager last-mile enrichment 上，生产者标记保持权威，Codex/OpenClaw 终止语义与现有 OTLP 对齐，失败 fail-open，且不再参与 checkpoint。未发现阻断问题，可以合入。3 条 Low：schema 未列出 gen_ai.turn.start/end、terminal 常量与 e2e validator 重复、默认 Agent 矩阵缺 dsh。

✅ Medium/High 及以上问题已全部解决，本轮评审通过（approved）。

Highlights（正向实践）

• inspect-before-fill，含生产者 false 也不覆盖
• 有界 process-local Map，明确不进 StateStore
• OpenClaw llm_output / Codex turn_status 特例有单测 + 集成测试

---

评审报告详见: code-review/pr-262/final-report.md
Generated by LoongSuite-Pilot Code Review Agent

[linrunqi08]

APPROVE-READY: no open Critical/High/Medium findings. Low items are follow-ups only.

Generated by LoongSuite-Pilot Code Review Agent