feat(eval): findings diagnostic chain with dual entry points

[MeloMei]

Eval 阶段三：把描述性数字升级为带证据的结论（finding），并让结论在 Skills 页被撞见。

做了什么

• 新增纯函数规则求值器（src/core/skill-eval-findings.ts），四条摩擦规则：已装未用 / 工具反复失败 / 触发轮次成本偏高 / 触发轮次频繁出错——每条规则标注严重度（low/medium，不设 high）、证据强度、样本量，样本不足即沉默
• 新增 listSkillToolOutcomes 查询：按工具名聚合 linked-turn 触发 turn 内的 kind='tool' span，只计 status='failed'（aborted 是用户中断不算失败），session_id 白名单从一开始就带 codex
• Eval 页实况卡片上方新增 Findings 段：每条 finding 可展开查看 observation + repairDirection + evidence，零 finding 时明示"不等于没问题"
• Skills 页技能列表新增 finding 数量徽标，点击跳转 Eval 页对应 Skill（双向入口），Eval 关闭时不渲染任何徽标
• 措辞纪律进测试：observation 不得出现因果词，repairDirection 只能是"Consider checking..."

边界

• V1 零改动，不新增表不新增迁移
• 不做因果断言、不设 high 严重度、findings 不落库（查询期计算）
• Qoder 会话在全部摩擦规则上沉默（无 trace_spans/token），显示 unobserved

测试

• 29 个阈值表驱动测试覆盖四条规则的边界（恰好命中 / 差一个样本 / 差一个百分点）
• V2 全量 292 文件 / 2526 passed，V1 97 文件 / 1235 passed，release-note:check 通过

[Copilot]

Pull request overview

This PR adds “findings” as a higher-level diagnostic layer in the Eval experience: it evaluates a small set of table-driven friction rules against observed skill usage/performance data, exposes them via IPC, and surfaces them both in the Eval page and as a badge entry point from the Skills page.

Changes:

• Introduces a pure-function findings evaluator (4 rules) plus wording-discipline tests to keep observations non-causal and repair directions in a constrained format.
• Adds IPC/service/store plumbing to compute per-skill findings and aggregated finding counts for Skills page badges.
• Updates the renderer UI/CSS to display findings in Eval and show clickable findings-count badges in the Skills list that deep-link into Eval.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
apps/main-2.0/src/shared/ipc/skills.ts	Adds IPC endpoints for findings + finding-counts.
apps/main-2.0/src/renderer/src/styles/skills-page.css	Styles the Skills-page findings badge link.
apps/main-2.0/src/renderer/src/styles/eval.css	Adds styling for the new Findings card UI.
apps/main-2.0/src/renderer/src/features/skills/skills-page.tsx	Threads findings-count data + navigation callback into the Skills list.
apps/main-2.0/src/renderer/src/features/skills/skill-library-list.tsx	Renders findings badge per skill with click-to-navigate behavior.
apps/main-2.0/src/renderer/src/features/eval/eval-page.tsx	Fetches and renders findings per selected skill; consumes preselected-skill navigation.
apps/main-2.0/src/renderer/src/App.tsx	Fetches global finding counts when Eval is enabled; wires Skills→Eval deep link.
apps/main-2.0/src/preload/skills.ts	Exposes findings IPC methods to the renderer API.
apps/main-2.0/src/main/services/skill-service.ts	Implements getSkillFindings + getSkillFindingCounts and calls the evaluator.
apps/main-2.0/src/main/services/skill-service.test.ts	Updates service harness mocks for the new store method.
apps/main-2.0/src/main/ipc/skills.ts	Registers IPC handlers for findings and finding counts.
apps/main-2.0/src/core/skill-eval-findings.ts	Adds table-driven findings evaluator + wording assertion helper.
apps/main-2.0/src/core/skill-eval-findings.test.ts	Adds boundary + wording-discipline tests for the 4 rules.
apps/main-2.0/src/core/session-store.ts	Adds session-store wrapper for listing tool outcomes.
apps/main-2.0/src/core/postgres/skill-repository.ts	Adds SkillToolOutcome model + query to aggregate tool-call outcomes per skill.
.release-notes/eval-skill-findings.md	Adds user-facing release note for Findings + dual entry points.

Suppressed comments (2)

apps/main-2.0/src/main/services/skill-service.ts:820

• getSkillFindingCounts uses linkedTriggers > 0 to decide whether a skill is "exercised" vs "never-used", which can incorrectly label triggered skills (with no linked triggers) as "never-used" and incorrectly count the installed-never-exercised finding. It also currently can’t represent "unobserved" for agents like trae, which can make the badge counts disagree with the Eval overview semantics.

      const installed = installedNames.has(name);
      const overviewLike = overviewItem
        ? {
            observation: (overviewItem.linkedTriggers > 0 ? "exercised" : "never-used") as "exercised" | "never-used" | "unobserved",
            installed,

apps/main-2.0/src/main/services/skill-service.ts:812

• getSkillFindingCounts runs getSkillPerformanceSignals + listSkillToolOutcomes inside a for ... of loop, resulting in 2×N round-trips (and serialized execution) for N skills. Since this is used to render a badge list, it can become noticeably slow with many installed/triggered skills.

    for (const name of allNames) {
      const [signals, toolOutcomes] = await Promise.all([
        store.getSkillPerformanceSignals(name),
        store.listSkillToolOutcomes(name),
      ]);

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.