feat(ai): admin view-transcript UI for crew/agent runs (AI-045)

CHANGELOG.md

@@ -2,6 +2,15 @@
 
 ## [Unreleased]
 
+### Phase 7 — admin "view transcript" UI for crew/agent runs (AI-045) (2026-06-15)
+
+Makes the multi-agent reasoning chain inspectable. Every crew run (`crew.autopublish`/`crew.seo`) and single-agent run (`studybuddy`) already persists an `agent_run` row with a nested-step transcript (researcher→drafter→critic→editor); AI-045 surfaces it in the admin app. **No schema change, no new table** — read-only over the existing Phase 6 `agent_run`.
+
+- **Backend** (`Api/Endpoints/AdminAiQualityEndpoints.cs`, mirrors the existing `GetTraces`/`GetTrace`): `GET /admin/ai-quality/agent-runs?agent=&limit=&offset=` (list, newest-first, `agent` filter = exact-or-prefix so `crew.` narrows to all crew runs; clamp 1–100; **list projection omits the heavy `StepsJson`+`Output`** and truncates `Goal` to 120 chars — guarded `Length > 120 ? Goal[..120] : Goal`) and `GET /admin/ai-quality/agent-runs/{id}` (detail, full **raw** `StepsJson` + `Output`, 404 on unknown). DTOs in `Contracts/Admin/AiQualityDtos.cs` (`AgentRunListItemDto`/`AgentRunsPageDto`/`AgentRunDetailDto`). `StepsJson` is passed through RAW and parsed client-side — same pattern as `TraceDetailDto`'s `MessagesJson`/`ToolCallsJson`; no brittle second server-side schema.
+- **Frontend** (`apps/admin/src/pages/AiQualityPage.tsx`, new `Transcripts` tab modeled on `TracesTab`/`TraceModal`): filterable list (All / `crew.autopublish` / `crew.seo` / `studybuddy`) + pager → click row → modal. The modal parses `stepsJson` into the step tree: each `sub_agent` step is a collapsible `{stage} · {agentName}` panel with its per-step usage; the **critic** panel is special-cased + default-expanded — its inner `llm_response` JSON is rendered as score chips (factual_accuracy/tone/length/banned_phrases) + a severity-colored issue list (blocker red / major amber / minor gray). Single-agent steps (`llm_response`/`tool_result`) render via `pretty()`. Every layer is defensive — a malformed/empty/non-JSON step falls back to raw and never throws.
+- **The JSON casing contract is the load-bearing risk** (a mismatch = blank transcript). The writer `DbAgentRunWriter` serializes `run.Steps` with **default** STJ options (unaffected by the HTTP pipeline's web-defaults), so the shape is mixed: top-level `AgentStep` records → PascalCase (`Index`/`Kind`/`Payload`/`At`), the `sub_agent` payload anon-object → camelCase (`stage`/`agentName`/`status`/`usage`/`steps`), the nested `AgentUsage` record → PascalCase again, inner `llm_response` payload `new { text }` → camelCase. The frontend reads each exactly. **`CrewTranscriptJsonContractTests`** (pure, no DB) serializes a real 4-stage `crew.autopublish` run through the same factory+options and asserts every key the UI depends on, **with negative asserts** so the test fails loudly if anyone ever puts camelCase/web options on the writer.
+- Entry points from the AutoPublish/SEO pages (deep-link by the runIds those `crew-generate` endpoints already return) are a deliberate fast-follow; this PR is the self-contained AI-quality tab. Admin-only (`/admin/*` auth) — the `Goal` can carry user passages / book source, acceptable for the owner's own audit surface.
+
 ### Phase 7 — synthetic-defect critic harness (AI-044) (2026-06-15)
 
 A calibration gate for the AI-041 `CriticAgent`: instead of trusting that the critic *would* catch a bad draft, we inject KNOWN defects into clean drafts and measure whether it actually does. ~23 fixtures over a single edition-description `ContentBrief` (`AutoPublishBriefs.Description("en")` — real 800–1600 char bounds + the shared `CrewBannedPhrases` blocklist): factual_hallucination ×6, banned_phrase ×4, length over/under ×2+2, tone_break ×4, plus 5 clean controls. The harness mirrors `ToolCallEvalRunner` exactly — real nano per case, **pure deterministic scoring, no judge**, `JudgeModelId="n/a"`, persists a reused `EvalRun` (**no schema change**), Score = catch-rate.

apps/admin/src/api/client.ts

@@ -485,6 +485,42 @@ export interface TraceDetail {
   userId: string | null
   createdAt: string
 }
+export interface AgentRunListItem {
+  id: string
+  agent: string
+  userId?: string | null
+  editionId?: string | null
+  status: string
+  goal: string
+  iterations: number
+  tokensIn: number
+  tokensOut: number
+  costUsd: number
+  latencyMs: number
+  hasError: boolean
+  createdAt: string
+}
+export interface AgentRunsPage {
+  total: number
+  items: AgentRunListItem[]
+}
+export interface AgentRunDetail {
+  id: string
+  agent: string
+  userId?: string | null
+  editionId?: string | null
+  status: string
+  goal: string
+  output?: string | null
+  stepsJson: string
+  iterations: number
+  tokensIn: number
+  tokensOut: number
+  costUsd: number
+  latencyMs: number
+  error?: string | null
+  createdAt: string
+}
 export interface EvalRun {
   id: string
   feature: string
@@ -1107,6 +1143,19 @@ export const adminApi = {
     return fetchJson<TraceDetail>(`/admin/ai-quality/traces/${id}`)
   },
 
+  getAgentRuns: async (params?: { agent?: string; limit?: number; offset?: number }): Promise<AgentRunsPage> => {
+    const query = new URLSearchParams()
+    if (params?.agent) query.set('agent', params.agent)
+    if (params?.limit) query.set('limit', String(params.limit))
+    if (params?.offset) query.set('offset', String(params.offset))
+    const qs = query.toString()
+    return fetchJson<AgentRunsPage>(`/admin/ai-quality/agent-runs${qs ? `?${qs}` : ''}`)
+  },
+
+  getAgentRun: async (id: string): Promise<AgentRunDetail> => {
+    return fetchJson<AgentRunDetail>(`/admin/ai-quality/agent-runs/${id}`)
+  },
+
   getAiEvals: async (params?: { feature?: string; limit?: number }): Promise<EvalRun[]> => {
     const query = new URLSearchParams()
     if (params?.feature) query.set('feature', params.feature)