Commit Graph

55 Commits

Author SHA1 Message Date
sinohqb
192fe0fc5f feat(safety): 扩展规则覆盖幻觉、越权和合规检查
All checks were successful
CI / test (pull_request) Successful in 4m2s
扩展 safety 规则,新增三个安全检查维度。

- 幻觉检测(check_hallucination):LLM 判断回答是否编造事实
- 越权操作检测(unauthorized_actions):模式匹配检测越权行为
- 合规性检查(required_disclaimers):检查必需声明是否存在
- 保持原有 blacklist + moderation API 向后兼容
- RuleResult.details 包含所有问题列表
- 新增 10 项单元测试(656 tests passed)

Closes #22
2026-08-25 15:30:00 +08:00
2f08e7bf06 Merge pull request 'feat(report): 上线验收报告 go/no-go 结论' (#28) from feat/go-no-go-verdict into main
All checks were successful
CI / test (push) Successful in 4m6s
2026-08-25 06:31:42 +00:00
sinohqb
2314ebe3bc feat(report): 上线验收报告 go/no-go 结论
All checks were successful
CI / test (pull_request) Successful in 4m7s
新增纯函数模块 evaluation/go_no_go.py,基于 RunSummary 指标自动生成
go/no-go/conditional 三级结论。

- AcceptanceCriteria 模型:judged_pass_rate_min、pass_rate_min、avg_latency_max_ms
- GoNoGoVerdict 模型:decision、summary、criteria_results
- evaluate_go_no_go() 纯函数:接收 summary dict + criteria -> 返回 verdict
- 集成到 generate_report(),报告 dict 自动附带 go_no_go 字段
- 三级结论:go(全部达标)、no_go(核心未达标)、conditional(有风险)
- 新增 9 项单元测试(642 tests passed)

Closes #20
2026-08-25 14:28:55 +08:00
sinohqb
a29f78ff4b feat(llm_score): 支持多维度独立评分
All checks were successful
CI / test (pull_request) Successful in 4m9s
扩展 llm_score 规则,支持通过 dimensions 参数配置多个评分维度,
每个维度独立评分 0-10 分。

- RuleResult 新增 details 字段存储结构化多维度分数
- 向后兼容:原有 criteria 单维度模式继续有效
- 多维度模式下各维度并行调用 LLM,返回平均分和明细
- 新增 4 项单元测试

Closes #19
2026-08-25 14:17:49 +08:00
sinohqb
bf1ec16ef6 perf(eval): case/rule/campaign 并发执行
- EvalEngine case 循环改 asyncio.gather + semaphore(默认 3 并发)
- 规则评估并行(默认 5 并发),LLM 评分耗时从串行求和降为最慢一条
- _resolve_model 改 async + 双重检查锁,保护共享模型缓存
- _case_errors / case_outcomes 并发写入加状态锁
- Campaign occurrence 派生并行(默认 2 并发),claim 拒绝时提前收敛
- 新增配置:max_concurrent_cases=3 / max_concurrent_rules=5 / max_concurrent_runs=2

SQLite StaticPool 单连接下 DB 写入仍天然串行,并行收益集中在
channel I/O 与 LLM 调用的等待重叠。
2026-08-24 23:18:11 +08:00
sinohqb
7eae6de52d refactor(evaluation/storage): 结算统一与 repository 拆分(Phase 2 + 3)
合并两个不可分割的深化:

Phase 2 — 智能作业结算统一(ADR-0012)
- intelligence_jobs.execute(job_kind, campaign_id, ...) 作为结算的
  唯一实现:建行 → 认领 → 校验 → generating → 落账,一处编排、
  一处截断(500 字符)。两个 executor 退化为 ensure_queued /
  validate / work_fn 三个小 adapter。
- analysis.validate_analysis_request() 共享校验入口(活动终态 →
  模型),路由捕获映射 400、executor 捕获落 failed 行,与
  validate_comparison_request 先例同构。
- campaign_runner._auto_start_analysis 的跳过守卫收敛至
  auto_intelligence_eligible 单一判断点。
- comparison.py 删除零调用的 build_comparison_payload;
  load_comparison_view 投影归位至 campaign_read_model。
- 新增 characterization 测试(认领竞争、重复触发、截断、恢复上限)。

Phase 3 — storage/repository.py 拆分
- AsyncJobRepository 及两个子类迁至
  storage/async_job_repository.py(Phase 2 的 intelligence_jobs
  与 comparison 必须 import 自该路径,故与 Phase 2 同 commit)。
- ExplorationSession / ExplorationMessage 迁至
  storage/exploration_repository.py;repository.py 由 1180 行降至
  约 814 行,grep 确认无残留符号。
- exploration 子模块与路由 import 全部更新;测试 import 跟随。

刻意不做:CAS 共享原语、app.py 五 registry 关停顺序归一
(ADR-0006 精神,等真实需求出现再议)。
2026-08-24 05:50:27 +08:00
sinohqb
1782b245bf refactor(architecture): deepen campaign runtime modules 2026-08-11 13:18:48 +08:00
sinohqb
864ae2b7fe fix(review): address release correctness findings 2026-08-09 03:20:40 +08:00
sinohqb
62485684ca fix(architecture): enforce lifecycle consistency 2026-08-07 10:59:27 +08:00
sinohqb
c896ab3f71 refactor(architecture): deepen evaluation lifecycle and read model 2026-08-07 03:11:37 +08:00
sinohqb
160332665e refactor(exploration): absorb settlement.py into ExplorationSessionRepository
Some checks failed
CI / test (push) Failing after 12s
将 settlement.py 的 settle_campaign_sessions 函数吸收为
ExplorationSessionRepository.expire_running_sessions 方法。删除浅模块
settlement.py(30 行,接口宽如实现),会话生命周期操作集中在 repository。

- 新增 ExplorationSessionRepository.expire_running_sessions(campaign_id)
- 更新 campaigns.py 和 campaign_runner.py 两个调用点
- 删除 backend/agenteval/exploration/settlement.py
- 所有测试通过,行为不变
2026-08-04 13:28:23 +08:00
sinohqb
c24998c762 refactor(metrics): extract dashboard aggregation to compute_dashboard
仪表盘聚合逻辑从 stats.py router 下沉到 metrics.py 的 compute_dashboard
纯函数。_settled 重命名为 settled_runs 并公开,_ts 重命名为 _sortable_ts。
router 从 40 行聚合逻辑缩到 5 行,只负责数据获取和序列化。

- 新增 compute_dashboard(runs, scenario_names, target_names) -> dict
- 新增 settled_runs(runs) 公开接口(原 _settled)
- trend 端点同步迁移到 settled_runs
- 5 个新测试覆盖 dashboard 聚合逻辑
2026-08-04 11:36:55 +08:00
sinohqb
42be31dd1f feat(report): add load_campaign_view as unified campaign read model
活动级读模型单一出口:一次取齐报告 / 探索 / 分析 / 对比四大数据源。
markdown handler 从 30 行拼装逻辑缩到 3 行;分析执行器同步迁移。
comparison.py 内部的 8 次 load_campaign_report 调用暂不动(跨请求冗余,
缓存收益有限,改动风险高)。

- 新增 load_campaign_view(session, campaign) -> dict[str, Any]
- 返回 {report, exploration, analysis, comparison} 四键
- 迁移 markdown handler 和分析执行器两个调用点
- 4 个新测试覆盖 view 的组装逻辑
2026-08-04 11:33:45 +08:00
sinohqb
2fddce8c92 refactor(case-verdict): extract build_case_evidence as single evidence-construction seam
用例判定证据构建收敛到 case_verdict.py 的 build_case_evidence 纯函数,
report.py 和 runs.py 各删 ~15 行重复逻辑,换一行调用。locality 回归:
证据构建改一处,全局生效。

- 新增 build_case_evidence(turns, results) -> dict[str, CaseEvidence]
- report.py:76-83 证据构建替换为一行调用
- runs.py:175-201 证据构建替换为一行调用
- 5 个新测试覆盖 build_case_evidence(纯函数,无 DB 依赖)
2026-08-04 11:23:56 +08:00
sinohqb
2b6cab6cb2 feat(comparison): unify read model and validation for period comparison
周期对比读模型升位为单一出口(load_comparison_view),GET/POST/markdown
三处调用点统一走同一 view 投影,消除「取数三件套」重复。校验逻辑收敛到
validate_comparison_request,router 捕获映射 400,执行器捕获落 failed 行,
校验顺序权威不再漂移。

- 新增 load_comparison_view:无行返回 status=none + auto_baseline,有行
  返回完整 comparison dict(含 model_name 标签)+ metric_diff
- 新增 validate_comparison_request:活动终态 → 模型 → 基线 → 分析,违
  规抛 ComparisonError
- execute_campaign_comparison 内联校验替换为 validate_comparison_request
  调用,catch ComparisonError 落 failed 行
- router 三处迁移:GET /comparison、POST /comparison、markdown 导出
- 删除 build_comparison_payload(已吸收进 load_comparison_view)
- 8 个新测试覆盖读模型三态 + 校验五错
2026-08-04 10:49:14 +08:00
sinohqb
f3a528611e refactor(tasks): route LLM background tasks through TaskRegistry
架构保养第二轮候选 1:分析 / 周期对比 / judge 复核三条 LLM 任务链
收进各自的模块级 TaskRegistry(强引用防 GC、按 id 幂等、shutdown
统一收敛),删除 judge 的 _BACKGROUND_TASKS 私货,start_* 不再返回
无人消费的 Task。启动清理块补两笔 orphan 清扫:滞留的 generating
分析与对比行标记为 failed,与僵尸运行清扫同构。新增 7 个单测。
2026-08-04 09:58:20 +08:00
sinohqb
38849d46f1 refactor(repository): narrow atomic updates for patrol/cancel/scheduler writes
全行 update 逼着调用方承诺「持有最新完整快照」,巡检水位、活动取消、
调度循环三方并发写入会互相覆盖(后写者抹掉先写者的 summary/status)。
CampaignRepository 新增三个窄口径原子操作:touch_patrol_watermark、
mark_cancelled、save_scheduler_state,各自只写自己的列;三处调用点迁移。
2026-08-04 03:39:42 +08:00
sinohqb
f8d8450b1e refactor(report): unify campaign report loading behind one read model
「活动报告取数三件套」此前在报告/markdown/分析/对比等 7 处手写重复,
唯一深化产物 build_campaign_report_dict 被锁在周期对比私有角落。
升位为 report.py 的 load_campaign_report(session, campaign) 单一出口
(探索线 summarize_campaign_exploration 同口径),并把 8 处
scenario_names 推导式收敛为 ScenarioRepository.name_map() 窄方法。
纯结构重排、零行为变更,572 项测试全绿。
2026-08-04 03:30:29 +08:00
sinohqb
ef4c094082 refactor(exploration): share one fetch+aggregate helper across outlets
report, markdown and analysis each repeated the
summarize_exploration(repo.list_by_campaign(...)) shape; collapse it
into summarize_campaign_exploration so the aggregation has one home.
2026-08-04 02:48:59 +08:00
sinohqb
62f7031b5b feat(exploration): expire dangling sessions on campaign finalize 2026-08-04 01:23:37 +08:00
sinohqb
2484c207af feat(exploration): findings flow into report, analysis and export
Exploration sessions aggregate into a single exploration summary
(session counts, goal-achievement rate, issue lists from experience
records, judge conclusions when reviewed) that feeds three exits:
the campaign report gains an exploration dimension, the v0.7 analysis
stage-two input gains the summary (stats only, never full dialogues),
and the Markdown export appends a findings appendix after analysis and
comparison. With no exploration data every output stays unchanged.
2026-08-03 19:16:33 +08:00
sinohqb
14b09e1ac6 feat(campaigns): 周期对比纳入 Markdown 导出,活动导出排版重优化
- 导出端点解析评测对象名与 completed 周期对比行(基线元信息、分析模型、现算机械 diff),渲染 `## 周期对比` 附录(趋势 + 指标变化表 + 问题演变 + 建议落实),紧跟智能分析之后;非 completed 则完全无痕
- 头部排版重优化:状态中文化、窗口与时段人类可读(24h、0h–1h)、友好时间戳、头部补评测对象名、「正式线」/「加速调试线 ×N」措辞(加速线附注压缩后实际耗时);Run 级导出不动
- 测试:渲染器黄金断言更新 + 附录/头部/缺省用例,集成测试新增导出含对比、无对比行、failed 行三例
2026-08-03 15:13:11 +08:00
sinohqb
dd3b9a5e91 refactor(comparison): 评审修复 — 共享 gateway_chat_client、指标元表、对比区块组件化
- analysis/comparison 重复的 _gateway_chat_client 提取为共享 gateway_chat_client
- Campaigns.tsx 周期对比区块抽为 PeriodComparisonSection 组件,指标格式化
  收敛为单一 METRICS 元表(消除三处 metric 分支级联)
- 基线下拉排除无 completed_at 的终态活动(选中必 400)
- spec/issue 03 追认 delta 按好坏着色口径与 GET 生效基线合并口径
2026-08-03 14:26:20 +08:00
sinohqb
1e55a21649 feat(comparison): v0.8 周期对比 — 计划指纹自动基线配对、机械指标 diff 与 LLM 演进叙述
正式线活动分析完成后自动链式生成对比;报告抽屉新增「周期对比」区块
(趋势徽章、指标 delta 表、问题演变、建议追踪,无自动基线时手动选择)。
版本号升至 0.8.0。
2026-08-03 13:32:48 +08:00
sinohqb
5d04455664 feat(campaigns): auto-trigger analysis and include it in markdown export
The scheduler loop enqueues the analysis task when a realtime campaign
completes; accelerated or cancelled campaigns and a missing analysis
model skip silently. The campaign markdown export appends the analysis
appendix (overall, problems, narratives, suggestions) when a completed
analysis exists.
2026-08-03 02:25:07 +08:00
sinohqb
15c542d92c feat(analysis): two-phase campaign analysis agent with storage and API
Add the analysis role's execution path: a two-phase orchestration
(per-scenario diagnosis gathered in parallel, then a synthesis pass)
that reads the existing campaign report aggregation plus capped failure
samples, validates the LLM's JSON against the report schema, and strips
fabricated run/scenario references before persisting. Results upsert one
row per campaign (generating/completed/failed) with the model config
snapshot; GET/POST /api/campaigns/{id}/analysis expose the state machine,
guarding non-terminal campaigns and missing analysis models with 400s.
2026-08-03 02:06:29 +08:00
sinohqb
8bc5aa6979 feat(campaign): add per-Run timeline seam + endpoint
build_campaign_timeline flattens a campaign's child Runs into offset-sorted
per-Run entries (distinct from the report's 12-bucket aggregation), reusing a
shared _run_window_offset口径 so both views place a run identically. Exposes
GET /campaigns/{id}/timeline and the api.ts type/call. (v0.6 ticket 06)
2026-07-31 16:49:04 +08:00
sinohqb
aa40c8e0d8 refactor(campaign): type Campaign.summary as CampaignSummary VO
Some checks failed
CI / test (push) Failing after 12s
Campaign.summary was a bare Optional[dict] while RunSummary is a typed VO —
scheduler state (spawned_indices/errors) flowed untyped through
set_/get_summary. Introduce CampaignSummary + SchedulerState (extra=allow,
validate_assignment), mirroring RunSummary; campaign_runner reads/writes the
VO. Also converge the ~10 repeated JSON column get/set pairs onto
_json_dumps/_json_loads helpers, unifying ensure_ascii=False and fixing the
set_modalities ensure_ascii=True trap.
2026-07-31 15:09:10 +08:00
sinohqb
ffc058f951 refactor(engine): thin _run_case and _save_rule_results
Three near-identical Turn(...)+save_turn blocks (send-fail / poll-except /
happy path) collapse into one _persist_turn helper differing only by the
optional fields set. The期望→隐式规则 translation becomes a pure,
unit-testable derive_implicit_rules seam (CONTEXT: 期望与规则叠加生效),
and the rule_type→ModelPurpose map is hoisted to a module constant.
2026-07-31 14:59:50 +08:00
sinohqb
9c01afa79b refactor(engine): extract build_run_summary pure seam
Single-run summary口径 (pass_rate / judged_pass_rate / avg_latency /
connectivity split) was inlined in run(), reachable only by driving a
whole async run, and report.py recomputed judged_pass_rate independently.
Extract build_run_summary — a pure function parallel to aggregate_runs
(cross-run) and combine_case_outcome (case-level). run() now collects
material and delegates; judged_pass_rate is stored in RunSummary so the
report reads it instead of recomputing.
2026-07-31 14:20:51 +08:00
sinohqb
983a58d013 refactor(verdict): unify read path on authoritative case_outcomes
Read paths recomputed per-case pass/connectivity independently — report
generation, the logs endpoint, and the frontend each derived it, and the
frontend's every(passed) recompute ignored the engine's authoritative
verdict. Extract resolve_case_verdicts: a single pure seam that prefers
stored case_outcomes verbatim and approximates only for legacy runs. The
logs endpoint now surfaces case_verdicts so the frontend reads instead of
recomputing.
2026-07-31 14:11:58 +08:00
sinohqb
f285738f6d refactor(report): split report generation from pure rendering
report.py mixed DB-reading generation with string formatting: the four
render_*_report(run_id, session) functions each re-fetched via
generate_report, so the HTML/Markdown/JSON formatting was welded to storage
and could not be unit-tested from a plain dict. Extract the formatting into a
new pure report_render module whose renderers take the already-built report
dict (no session, no storage import). Migrate every caller to generate-then-
render, delete the old coupled renderers with no back-compat shim, and drop
the _aggregate_runs middle-man alias in favour of metrics.aggregate_runs.
2026-07-31 10:19:04 +08:00
sinohqb
0cca4963d1 refactor(tasks): unify run/campaign task registries into TaskRegistry
Both the single-run path and the campaign scheduler drove long-lived
asyncio tasks through their own duplicated _tasks/_cancel_events dicts and
shutdown loops. Collapse them into one deep TaskRegistry module,
instantiated as run_registry and campaign_registry. launch() creates the
cancel event before the task (so a cancel during startup is never lost),
wires done-callback cleanup, and is idempotent per id; this makes runs.py's
hard-cancel fallback provably dead, so it is removed. App shutdown now
gracefully stops in-flight runs too, not just campaigns.
2026-07-31 03:39:03 +08:00
sinohqb
e815298ce5 refactor(campaign): move tick decisions into the pure scheduler seam
Extend the pure scheduler with elapsed_seconds (clock injected), decide_tick
(offset + due + lifecycle action) and resolve_finalize (cancel-race guard),
so the durable loop stops hand-coding elapsed/finished/status checks and only
does I/O. Deletes the runner's private _elapsed_seconds and converges
current_window_offset onto the one pure elapsed computation. The clock-skew
tolerance and cancel-race guard are now unit-testable at the seam.
2026-07-31 02:20:14 +08:00
sinohqb
782916a283 refactor(metrics): type Run summary and converge cross-run aggregation
Give EvalRun.summary a typed RunSummary value (unified RunError, lenient
legacy parsing) so readers stop reaching into a schemaless dict, and route
every cross-run rollup — dashboard, scenario ranking, trend, campaign
report — through one aggregate_runs seam. Fixes the divergence where
stats averaged pass_rate over completed-only runs while the campaign
report counted faults as 0.0. Cross-run rule (ADR-0004): genuine faults
count 0.0, user-cancelled runs are excluded from both denominators.
2026-07-31 01:57:56 +08:00
sinohqb
7ed765726f feat(campaigns): live list progress, polling, and richer drill-down
Embed compact progress (completed/planned total + overall pass_rate,
reusing the report's aggregation) into GET /campaigns so the list drops
its N+1 report fetch. Poll list and open report drawer every 5s while the
tab is active and a campaign is still running. Show scenario version and
trigger source tags in the child-run drill-down.
2026-07-30 15:35:18 +08:00
sinohqb
f433ebb970 feat(campaigns): dual-axis periodic report (time trend + capability)
Add generate_campaign_report: a pure aggregator over a campaign's child Runs
producing a time-trend axis (Runs bucketed by service-window position) and a
capability-summary axis (grouped by scenario), each carrying pass_rate /
availability / latency. pass_rate keeps the single-Run case-level meaning and
counts execution failures as 0.0 (ADR-0002); time_scale only places Runs into
window-time buckets and never alters any figure. Engine summary now records
avg_latency_ms to feed the latency axis.

Expose GET /api/campaigns/{id}/report (structured) and .../report/markdown
(reusing the existing Markdown export path). Adds "可用性/Availability" to the
domain glossary.
2026-07-30 13:55:32 +08:00
sinohqb
8910fd17e0 feat(campaigns): durable scheduler loop with restart recovery and cancel
Add a thin async loop (run_campaign_loop) that ticks on real wall-clock time,
maps elapsed×time_scale to a window offset via the pure decide_schedule, spawns
due child Runs, and marks the campaign COMPLETED at window end. All authority
lives in the DB (started_at, spawned_indices, status), so the app lifespan can
resume every RUNNING campaign on startup without double-spawning and stop all
loops gracefully on shutdown. A failing plan entry is skipped and recorded
rather than wedging the campaign.

Creating a campaign now starts its loop; POST /api/campaigns/{id}/cancel stops
further spawning (completed child Runs are kept); GET /api/campaigns/{id}
reports live progress (window offset, spawned/completed Run counts).
2026-07-30 13:33:10 +08:00
sinohqb
c6b102a9b5 feat(campaigns): add scheduling decision and child-Run spawning
Add the pure scheduling seam (campaign_scheduler.decide_schedule) that, given a
static plan and window-clock offset, decides which plan entries are due and
whether the window ended — mirroring judgement.combine_case_outcome, with
time_scale confined to the clock mapping so it never touches judgement/report.

The campaign_runner shell maps injected elapsed time to a window offset, spawns
due child Runs through the existing EvalEngine.run(existing_run=...) path with
campaign_id + RunTrigger.CAMPAIGN, and persists spawned-entry indices per entry
for idempotent, restart-recoverable progress. No auto loop yet (ticket 03).
2026-07-30 12:06:31 +08:00
sinohqb
1345daddd2 feat(engine): make poll_reply timeout configurable via env
Some checks failed
CI / test (push) Failing after 45s
被评数字员工响应普遍逼近 30s 硬编码轮询超时,越线的轮次被记为无回复
(run 427b14bb round 2 实测 31.4s 超时)。新增
AGENTEVAL_POLL_REPLY_TIMEOUT(默认 30s),engine 未显式传入
timeout_config 时从 settings 取值,慢目标可放宽。
2026-07-30 09:58:08 +08:00
sinohqb
5db0ede4f4 refactor(judgement): converge case-pass decision into one deep module
Some checks failed
CI / test (push) Failing after 50s
「用例是否通过」此前散落 8 处且互相矛盾:engine 权威判定焊死在持久化里
不可单测;report 聚合/compare/markdown 各自从规则结果反推,规则还不一致
(markdown 用 all([]) 把故障用例误渲染成 )。

- 新增纯函数 evaluation/judgement.combine_case_outcome(RuleOutcome/
  CaseOutcome),判定组合脱离通道与 DB 可单测(判定矩阵 14 例)
- engine 调用它一次,逐用例权威结果写入 summary.case_outcomes(JSON,
  零迁移);report/compare/markdown 只读权威值,老 run fallback 反推
- 故障用例判 False(ADR-0002):修正 markdown 的  bug 与 compare 的
  None;顺带修 engine 连通用例无回复也算通过的 bug
- pass_rate 口径改为用例级(CONTEXT.md 词条),规则级保留在
  passed_rules/total_rules;CLI 对比标签同步更正
- 修 RunRepository.update 漏拷 scenario_version/triggered_by 的字段漂移
2026-07-29 19:45:02 +08:00
sinohqb
fbac28bc7e feat(frontend): always show trigger source labels (手动/AI 助手/CLI)
Some checks failed
CI / test (push) Failing after 49s
运行列表、报告页运行下拉、仪表盘最近评测行不再隐藏手动标签;
报告头与对比 A/B 卡片新增"触发方式"(报告 payload 补 triggered_by)。
2026-07-29 14:34:32 +08:00
sinohqb
770d260750 feat(report): compare requires same scenario version (ticket 05)
Some checks failed
CI / test (push) Failing after 39s
对比报告可比性收紧为同场景同考纲版本(ADR-0001):跨版本 API 返回 400
(detail 含双方版本号),报告生成层抛 ValueError;前端对比候选按
同场景 + 同版本过滤,A 变更后自动清空不可比的 B。文档"尚未实现"标注移除。
2026-07-29 11:21:52 +08:00
sinohqb
0a47260237 feat(run): snapshot scenario version at run creation (ticket 04)
运行创建时快照场景考纲版本,三种触发来源(手动/AI 助手/CLI)一致;
迁移回填存量运行为其场景当前版本,孤儿运行回填 1。运行列表、
报告头与对比卡片展示 v{n} 版本标签。
2026-07-29 10:59:44 +08:00
sinohqb
8a526599ab feat(report): annotate connectivity cases and add judged pass rate (ticket 02)
报告层推导连通用例标记(无判定结果 + 每轮有回复 + 无用例级错误),
summary 新增 connectivity_cases 与 judged_pass_rate(无判定型用例时为 null)。
对比报告同步标注且连通用例按引擎口径计通过;总通过率口径不变(ADR-0002)。
2026-07-29 10:32:56 +08:00
sinohqb
5dd1bc8535 feat(engine): expectations now additive with explicit rules (ticket 01)
期望始终派生隐式判定并与显式规则叠加执行:rule_logic 只组合显式规则,
期望是叠加其上的硬约束,任一不满足即用例不通过。隐式判定以 EvalResult
同构落库,reason 前缀 [期望] 标明来源。连通用例(无规则无期望)行为不变。
2026-07-29 10:24:01 +08:00
sinohqb
739d586aec feat(backend): v0.4 triggered_by tracking, login gate, compare guard, dashboard stats
- EvalRun.triggered_by 全链路(manual/ai_assistant/cli)+ 迁移 b7d4e6f81c22
- 标准 agenteval-run SKILL.md 纳入版本管理,deploy 脚本同步 + API Key 注入
- 简单登录:AGENTEVAL_ADMIN_PASSWORD + HMAC 会话 token,require_auth 双凭据
- 对比报告限同场景(400)+ 空 results 误判修复
- /api/stats/dashboard 扩展聚合;/api/runs 返回场景/对象名
- 测试 218 → 232
2026-07-28 17:40:54 +08:00
sinohqb
470ff5875f feat(models): add centralized model configuration 2026-07-17 20:02:43 +08:00
sinohqb
867d4e3ff1 fix(engine): dynamic 生成失败原因持久化到 run.summary.case_errors
## 背景
用户反馈动态问诊评测「执行不下去」。诊断发现:dynamic 用例的 LLM 消息
生成 API 调用失败(0.2s 瞬间 failed,凭证/参数问题),引擎正确地标记
case 失败——但失败的具体原因(如 401 详情)只 emit 到 WebSocket,从不
写入 run.summary。导致 run 记录只有 total_rules:0 failed,DB/报告查不到
任何原因,用户和排查者都无从下手。

## 修复
- EvalEngine 新增 self._case_errors 收集致命的 case 级错误
- _generate_messages 的 8 个失败点统一走 _fail() helper:既 emit 到
  WebSocket,也记录到 _case_errors(含 case_id + stage + 具体 error)
- run() 汇总时把 _case_errors 写入 summary["case_errors"]
- 新增测试:dynamic 生成失败时 summary.case_errors 必须含原因(补上
  之前 KNOWN-2 记录的 _generate_messages 测试盲区)

## 注
这不是导致失败的 bug(失败源于外部 API 凭证/参数),而是让失败「可诊断」
的可用性修复。用户需自查 llm_config 的 api_key 是否有效/model 是否被
该端点接受。

Co-Authored-By: Claude <noreply@anthropic.com>
2026-07-17 16:29:51 +08:00
sinohqb
c4962ddadf fix(llm_score): 修复多轮用例 question 提取错位导致普遍打 0 分
## 现象
多轮/动态用例的 llm_score 规则几乎全部返回 0.0/10 失败,即便被测对象
回复质量很高(实测 400+ 字的专业医疗回复也是 0 分)。

## 根因
evaluate() 里 question 提取逻辑错误:
    if len(dialog) >= 2:
        question_text = extract_reply_text(dialog[-2].reply)  # BUG
dialog[-2].reply 是「上一轮智能体的回复」,被误当成「用户问题」。于是评分
LLM 收到的问答对是:
  - "用户问题" = 上一轮 AI 回复
  - "智能体回复" = 当前轮 AI 回复
两段都是 AI 说的话、互不相关,评分 LLM 判定牛头不对马嘴 → 打 0 分。

单轮用例因 len(dialog)<2 走 sent_message 提取(正确),故不受影响;
问题只在多轮/dynamic 用例爆发。

## 修复
- question 始终取当前轮 last_turn.sent_message(用户实际发送的问题),
  移除错误的 dialog[-2].reply 分支
- reason 增加评分 LLM 自己的理由(parsed["reason"]),便于未来诊断
- 新增回归测试:多轮场景验证 question 来自 last sent_message 而非 prev reply

Co-Authored-By: Claude <noreply@anthropic.com>
2026-07-17 16:15:07 +08:00