fix(wukong): rework event-log conversion for correct GenAI traces

[Snssn]

Summary

Wukong sessions produced malformed GenAI traces. This reworks the AGUI → event-log conversion in WukongInput to fix four issues, all verified end-to-end against real wukong-cli data:

• Truncated answer / zero tokens — tighten the completeness gate: a message is only emitted when it has settled (isComplete + RUN_FINISHED/RUN_ERROR + USAGE + a closing TEXT_MESSAGE_END). A still-streaming message is deferred to the next poll instead of being emitted truncated, and the cursor is not advanced past it.
• Missing other event — each turn now begins with an other user-prompt event, so the stream no longer starts with tool.call.
• Wrong step order — steps are segmented by LLM decision (an assistant utterance plus the tools it triggers), USAGE is attributed per step, and all records are sorted by time_unix_nano (other → llm.request → llm.response → tool.call → tool.result → …).
• Token parsing — read real prompt/completion/total/cached tokens from USAGE.

Additional cleanups:

• Merge REASONING + text into a single llm.response (multi-part).
• Use a unique per-utterance response id (not the shared runId) so the converter does not collapse distinct steps.
• Drop redundant agent.* extension fields (source/message_id/conversation_id/ttft/e2e_ttft/run_duration) that duplicated standard fields.

Test plan

• npx vitest run tests/unit/inputs/wukong-input.test.ts — 37/37 pass
• New real-session fixture (tests/fixtures/wukong/leetcode-session.json) + end-to-end assertions (other-first, full answer, real tokens, STEP == LLM, chronological order)
• Verified on live wukong-cli data: field-coverage.mjs --agents wukong = 100% for all event types; validate-trace.mjs = 0 errors

[linrunqi08]

🔍 Code Review Summary

PR #170 fix(wukong): rework event-log conversion for correct GenAI traces · head 8474be05

Severity	Count
Critical	0
High	3
Medium	3
Low	2 (+ 若干既有问题)

Lifecycle Verdict

Check	Result
资源释放	PASS
死锁/卡死风险	PASS（graceful stop 因 abort 时序失效可阻塞 ~10s，既有）
状态恢复正确性	FAIL

状态恢复 FAIL 证据：

• H2：已 RUN_FINISHED、isComplete=1 但缺 USAGE 的消息会让 isMessageComplete 永久返回 false，findLastCompleteIndex 在此处 break 且游标不前进 → 该 session 从此条起所有后续 turn 永久静默不发射。
• 既有项：transformAssistantMessage 对某条消息抛异常被 try/catch 吞掉后，newSeenCount 仍按 processable.length 越过被丢消息，永久跳过。

总体结论

方向正确：other-first、答案不截断、真实 token、时间序、reasoning+text 合并单条 response、per-utterance response.id 等均与 EVENT_LOG_TO_TRACE_SPEC 一致，并在 leetcode 真实会话上验证。

但重写在删除既有能力/权威信号时引入回归，建议合入前修复 3 个 High：

1. [High] ACTIVITY_SNAPSHOT 抽取丢失（wukong-input.ts:920）：删除了 FILE_READ/SEARCH/SKILL/ARTIFACT 的专用抽取，落入 default 后 tool.call.arguments/tool.call.result 被完全丢弃。本 PR 自带 fixture 的 FILE_READ 活动即触发（读取的 URL/内容全丢），但 e2e 未断言 args/result 故 CI 全绿放行。
2. [High] isMessageComplete 过度门禁（wukong-input.ts:1099）：isComplete===1 未被视为权威，仍强制要求 USAGE 等 → 头阻塞/永久丢数据（见 Lifecycle）。相对 base 的回归。
3. [High] step 分段欠切分（wukong-input.ts:501）：弃用 STEP_STARTED，仅按发言边界切 step；两次"中间无 text/reasoning"的连续工具轮次会被并入一个 step → 违反 spec §2.3（STEP 数 ≠ 真实 LLM 调用数）。

Medium：activity 结果状态判定回归（失败被标 success、error.message 丢失，:949）；多条 USAGE 下 AGENT token 可能双算/少算（:570）；测试覆盖盲区（activity args/result、reasoning+text 多 part 均无断言）。

Highlights（正向实践）

• other-first + messages_delta 增量 + 按 time_unix_nano 排序，忠实落地 spec §5/§7/§8。
• reasoning+text 合并单条 llm.response、per-utterance 唯一 response.id，规避 §4.2 双 span 与 step 折叠。
• USAGE 落最终 step 使 AGENT 聚合 == 真实上报（单聚合场景），并有 Σ tokens 守恒的 e2e 断言。
• 外部数据防御到位：时间戳 sanitize、numOr 拒 NaN/Inf、tool args JSON.parse try/catch 回退。

---

评审报告详见: code-review/pr-170/final-report.md
Generated by LoongSuite-Pilot Code Review Agent

[Snssn]

感谢细致的 review 🙏 已在 59999a4 全部修复，逐条回应：

[High] #1 ACTIVITY_SNAPSHOT 抽取回退丢数据 — 已修复。恢复 FILE_READ / SEARCH / SKILL / ARTIFACT 的专用抽取，并补回 FILE_WRITE 的 file_path 回退与 DIRECTORY_LIST 的 files/total_count 回退，compactObject 一并恢复。default 分支不再吞掉这些类型的 args/result。

[Medium] #2 activity 结果状态判定回归 — 已修复。恢复 resolveActivityResultStatus（识别 status 字符串 error/failed/failure、非空 error_message、cancelled）并重新输出 error.message，不再只看 exit_code。

[High] #3 完成门禁覆盖 API 权威标志 — 已修复。isComplete===1 直接判定完成、===0 推迟，仅在该标志缺失时才回退到 RUN 终止启发式，避免单条消息永久阻塞整条 session。

[High] #4 step 分段欠切分 — 已修复。存在 STEP_STARTED/STEP_FINISHED 时按其权威分段；缺失时对"当前 step 已产出工具后、开始时间晚于上一个工具结束"的顺序工具开启新 step（重叠/并行工具仍归并同一 step，符合 §2.3）。

[Medium] #5 多条 USAGE 双算/少算 — 已改为按 step 累加（不再覆盖）。

[Medium] #6 测试盲区 — 已补：e2e 断言 FILE_READ 的 args(path)/result(content)；新增 FILE_READ/SEARCH/SKILL/ARTIFACT 载荷抽取、status-from-error_message、REASONING+TEXT 合并单条 llm.response 多 part 的用例。

另外修复了一个连带问题：isComplete 生效后，尾部的 PERMISSION（HITL 审批）活动消息与 invisible 的 session_context_stats 元数据消息会被处理并产生碎片 trace / 空 tool。现已跳过 PERMISSION 活动与无 run/无内容的辅助消息，保证单轮单 trace、顺序正确。

已用真实 wukong 会话复核：field-coverage --agents wukong 全字段 100%，validate-trace 0 error；单测 41/41 通过。

[ralf0131]

Review: fix(wukong): rework event-log conversion for correct GenAI traces

Summary

Focused 3-file refactor (−62 net lines) that reworks the AGUI → event-log conversion in WukongInput to fix malformed GenAI traces. Four issues addressed, all verified against real wukong-cli data.

Key Changes Analysis

1. Tightened completeness gate

• A message is only emitted when it has fully settled: isComplete + RUN_FINISHED/RUN_ERROR + USAGE + closing TEXT_MESSAGE_END
• Still-streaming messages are deferred to next poll — prevents truncated answers and zero-token emission
• Cursor is not advanced past incomplete messages — correct behavior

2. New StepAcc accumulator interface

• Properly segments one assistant utterance (reasoning + text) plus triggered tools
• Per-utterance responseId ensures unique gen_ai.response.id per LLM call (run-level runId is shared and must not be reused)
• Clear documentation: "one step == one LLM decision (spec §2.3)"

3. Step segmentation

• Uses explicit STEP_STARTED/STEP_FINISHED boundaries when available (authoritative)
• Falls back to heuristic segmentation when not present — good defensive design

4. Defensive timestamp sanitization

• evs map sanitizes AGUI timestamps once upfront with fallback to msg.createdAt
• Handles external data gracefully

5. Metadata-only message filtering

• MEANINGFUL_EVENTS set skips messages with only CUSTOM/FIRST_TOKEN events
• Prevents spurious empty LLM pairs — good edge case handling

6. other event per turn

• Each turn now begins with an other event for the user prompt
• Uses earliest user message timestamp for correct temporal ordering

Test Coverage

• wukong-input.test.ts significantly updated (+238 −177) — tests verify the new conversion logic
• New fixture leetcode-session.json (+80) — real test data for end-to-end verification
• CI all green: build-and-test (Node 18/20/22) ✅

Minor Observations

1. ACTIVITY_TYPE_TO_TOOL_NAME changes: Removed FILE_READ and SEARCH mappings. If these activity types still appear in real AGUI data, they'll fall through to the default case. Worth confirming these are truly unused.

2. stepMessageId format: Changed from (evt.messageId as string) ?? 'step-${stepIndex}' to just 's${stepIndex}'. This simplifies the ID but loses the AGUI messageId correlation. If downstream consumers relied on this correlation, it could break.

Verdict

Clean, well-documented refactor that fixes real trace correctness issues. Good test coverage with real fixtures. CI green. Approving.

---

Automated review by github-manager-bot

[linrunqi08]

🔍 Code Review Round 2 — 修复验证

PR #170 · head 59999a48（round-1 head 8474be05）· 增量验证 commit 59999a48

round-1 全部 6 条 inline 评论（3 High + 3 Medium）均已修复，逐条已回复并 resolve。作者另主动修复 PERMISSION/invisible 消息碎片 trace 连带问题，核对不引入回归。

项	级别	结论
#1 ACTIVITY_SNAPSHOT 抽取回退	High	✅ fixed
#2 结果状态判定回归	Medium	✅ fixed
#3 完成门禁覆盖 isComplete	High	✅ fixed
#4 step 分段欠切分	High	✅ fixed
#5 USAGE 双算/少算	Medium	✅ fixed
#6 测试覆盖盲区	Medium	✅ fixed

Lifecycle Verdict

Check	round-1	round-2
资源释放	PASS	PASS
死锁/卡死风险	PASS	PASS
状态恢复正确性	FAIL	PASS（H2 根因已修）

Merge Gate（合入门禁）

APPROVE-READY ✅ — 无任何 open 阻断项（Critical/High/Medium 全部 fixed，Lifecycle 三项全 PASS）。仅剩 2 项 Low（otherTs 未钳制、stepMessageId 死字段）不阻断合入。

验证

• npx vitest run tests/unit/inputs/wukong-input.test.ts → 41/41 pass（node v22）。

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

---

评审报告详见: code-review/pr-170/final-report-round2.md
Generated by LoongSuite-Pilot Code Review Agent

[ralf0131]

⚠️ Merge conflict detected

This PR has conflicts with the main branch and cannot be merged. Please rebase or merge main into your branch and resolve the conflicts:

git fetch origin
git checkout fix/wukong-trace-quality
git rebase origin/main
# resolve conflicts, then:
git push --force-with-lease

This is a one-time reminder. Feel free to @mention me for a re-review after conflicts are resolved.

---

Automated notification by github-manager-bot

[ralf0131]

Summary

Reworks the AGUI → event-log conversion in WukongInput to fix four issues with Wukong session GenAI traces: truncated answers/zero tokens, incorrect step segmentation, missing response IDs, and orphan user messages.

Overall: LGTM. The refactoring significantly improves code clarity and correctness.

Highlights

• StepAcc accumulator pattern — replaces the old mutable flushStepLlm closure pattern with a clean, explicit accumulator per step. Each step now collects reasoning, text, tool calls, and usage independently, making the code much easier to reason about
• Run-level / step-level separation — RUN_STARTED/RUN_FINISHED/RUN_ERROR are now scanned in a dedicated pre-pass, cleanly decoupled from step segmentation logic
• Completeness gate hardening — the new isComplete field from get_spark_agui_messages provides an explicit signal; fallback to event-based detection for older payloads maintains backward compatibility
• Metadata-only message filtering — MEANINGFUL_EVENTS set prevents spurious empty LLM pairs from invisible session_context_stats messages
• Per-utterance response IDs — each LLM call in a run now gets a unique gen_ai.response.id, fixing the previous issue where all steps shared the run-level ID
• User prompt timestamp — userPromptTs correctly timestamps the other event before the run starts

Minor Observations (non-blocking)

1. Tool name mapping cleanup — FILE_READ and SEARCH removed from ACTIVITY_TYPE_TO_TOOL_NAME. If any legacy Wukong data still emits these activity types, they will be silently dropped. Consider logging a warning for unmapped types in a follow-up
2. openStep timestamp — uses the triggering event timestamp, which is correct. If the event timestamp is somehow missing/invalid, the fallback to msg.createdAt in the event sanitization pass handles it

---

Automated review by github-manager-bot

[ralf0131]

Summary

Complete rework of Wukong event-log conversion for correct GenAI trace generation. Replaces the monolithic transformation with a clean step-based accumulator model (StepAcc) that correctly segments multi-step conversations.

Highlights

• Correct step segmentation: One step = one LLM decision (reasoning + text + tools). New step opens when fresh utterance arrives after tools emitted
• Completeness gate: isComplete flag + event-based fallback prevents truncated answers and 0-token emissions
• Defensive timestamp handling: Sanitizes timestamps up front, handles metadata-only messages correctly
• Per-utterance response IDs: Each LLM call gets a unique responseId (not shared run-level id)
• Clean separation: Run-level scan (RUN_STARTED/FINISHED/ERROR) separated from step-level accumulation

Comprehensive test coverage with real fixture data.

---

Automated review by github-manager-bot