@app-vitals/agent-admin 0.0.1 → 0.1.0

package/src/agent-cron-run-stats.ts

@@ -0,0 +1,752 @@
+/**
+ * admin/src/agent-cron-run-stats.ts
+ * AgentCronRunStatsService — aggregates AgentCronRun token columns into
+ * several dimensions: totals, byAgent, byCron (with AgentCronJob name JOIN),
+ * byModel, daily (DATE(startedAt)), byCronModel, and byPhase.
+ *
+ * Skipped runs (skipped=true) are always excluded from all aggregations.
+ * byPhase additionally excludes runs with a null phaseId (legacy five-job
+ * crons, or runs that predate phase tracking, or runs whose agent hasn't
+ * reconciled its phase child rows yet).
+ *
+ * bySkill (PAU-1.6) aggregates AgentCronRunSkillUsage per (kind, name) over
+ * non-skipped runs only. baselines groups the first-turn context baseline per
+ * (contextFingerprint, baselineModel, phase) — skipped runs ARE included
+ * (their baseline is still a valid measurement), runs with no baseline
+ * (baselineContextTokens NULL) are excluded; rows are ordered by first
+ * appearance. Both are empty arrays when nothing has been reported.
+ * Uses $queryRaw for all dimensions — Prisma groupBy doesn't support the
+ * LEFT JOIN needed for byCron.
+ *
+ * byCron/byCronModel additionally group by phase (WL-3.5, LPC-3.1): each row
+ * carries a `phase` field derived by LEFT JOINing "AgentCronJob" a second
+ * time (alias `p`) on `r."phaseId" = p.id` and stripping the
+ * "shipwright-" prefix from p.name (e.g. "shipwright-dev-task" → "dev-task")
+ * so the output label matches exactly what the old `r.phase` string column
+ * used to contain. A run with no phaseId (legacy five-job crons, or an
+ * unreconciled agent) has phase NULL, and since SQL GROUP BY treats NULL as
+ * a single group, all of a cron's legacy runs still collapse into one row —
+ * identical to today's cronId-only display. Runs that do carry a phaseId
+ * (unified loop) produce one row per (cronId, phase), surfacing the
+ * dev-task/review/patch/deploy breakdown that used to be implicit in having
+ * five separate crons.
+ */
+
+import type { PrismaClient } from "../prisma/client/client.ts";
+import { Prisma } from "../prisma/client/client.ts";
+
+// ─── Types (mirrored from metrics/src/lib/admin-metrics-client.ts) ───────────
+// These types are defined here to keep admin self-contained (rootDir constraint).
+// The shapes must stay in sync with the interfaces in admin-metrics-client.ts.
+
+interface TokenAggregate {
+  input: number;
+  output: number;
+  cacheRead: number;
+  cacheCreation: number;
+  total: number;
+  costUsd?: number;
+}
+
+interface KeyedTokenAggregate extends TokenAggregate {
+  key: string;
+}
+
+interface DoubleKeyedTokenAggregate extends TokenAggregate {
+  key1: string;
+  key2: string;
+  /**
+   * Pipeline phase this row's runs served (dev-task/review/patch/deploy).
+   * Only populated on byCron/byCronModel rows; null for legacy runs that
+   * predate phase tracking, and always undefined on other dimensions
+   * (byAgent, byModel) which don't group by phase.
+   */
+  phase?: string | null;
+}
+
+interface DailyTokenAggregate extends TokenAggregate {
+  period: string;
+}
+
+export interface CronRunTokenStats {
+  totals: TokenAggregate;
+  byAgent: KeyedTokenAggregate[];
+  byCron: DoubleKeyedTokenAggregate[];
+  byModel: DoubleKeyedTokenAggregate[];
+  daily: DailyTokenAggregate[];
+  byCronModel: DoubleKeyedTokenAggregate[]; // key1=agentId:cronName, key2=model
+  byPhase: KeyedTokenAggregate[]; // key=phase; runs with a null phase are excluded
+  bySkill?: SkillUsageAggregate[]; // per (kind,name); skipped runs excluded
+  baselines?: ContextBaselineAggregate[]; // per (fingerprint,model,phase); runs w/o baseline excluded
+}
+
+/** Per-skill / per-subagent usage rollup across runs. */
+interface SkillUsageAggregate {
+  kind: string;
+  name: string;
+  runs: number;
+  invocations: number;
+  turns: number;
+  input: number;
+  output: number;
+  cacheRead: number;
+  cacheCreation: number;
+  /** Mean invokeContextDelta over rows that reported one; null when none did. */
+  avgInvokeContextDelta: number | null;
+}
+
+/** First-turn context baseline rollup for runs sharing a fingerprint/model/phase. */
+interface ContextBaselineAggregate {
+  contextFingerprint: string | null;
+  baselineModel: string | null;
+  phase: string | null;
+  runs: number;
+  avgContextTokens: number;
+  minContextTokens: number;
+  maxContextTokens: number;
+  /** Mean over runs that reported a value; null when none did. */
+  avgTurns: number | null;
+  avgToolCalls: number | null;
+  firstSeen: string;
+  lastSeen: string;
+}
+
+interface CronRunOutcomeSeries {
+  /** Pipeline phase ("dev-task", "review", ...); null for runs without a phaseId. */
+  phase: string | null;
+  /** Context fingerprint stamped by the agent; null for runs that predate it. */
+  contextFingerprint: string | null;
+  runs: number;
+  completed: number;
+  failed: number;
+  skipped: number;
+  /** skipReason → count, over this series' skipped runs (null reasons key as "unknown"). */
+  skipReasons: Record<string, number>;
+  /** Over non-skipped runs with a completedAt; null when there are none. */
+  avgDurationMs: number | null;
+  p50DurationMs: number | null;
+  /** Over non-skipped runs that reported the metric; null when none did. */
+  avgTurns: number | null;
+  avgToolCalls: number | null;
+  avgContextTokens: number | null;
+}
+
+export interface CronRunOutcomes {
+  series: CronRunOutcomeSeries[];
+}
+
+interface OutcomeRow {
+  phase: string | null;
+  context_fingerprint: string | null;
+  runs: bigint;
+  completed: bigint;
+  failed: bigint;
+  skipped: bigint;
+  avg_duration_ms: number | null;
+  p50_duration_ms: number | null;
+  avg_turns: number | null;
+  avg_tool_calls: number | null;
+  avg_context_tokens: number | null;
+}
+
+interface SkipReasonRow {
+  phase: string | null;
+  context_fingerprint: string | null;
+  skip_reason: string | null;
+  n: bigint;
+}
+
+// ─── Raw row types from $queryRaw ────────────────────────────────────────────
+
+interface TotalsRow {
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd: number | null;
+}
+
+interface ByAgentRow {
+  agent_id: string;
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd: number | null;
+}
+
+interface ByCronRow {
+  agent_id: string;
+  cron_id: string;
+  cron_name: string | null;
+  phase: string | null;
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd: number | null;
+}
+
+interface ByModelRow {
+  agent_id: string;
+  model: string;
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd: number | null;
+}
+
+interface DailyRow {
+  period: string;
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd: number | null;
+}
+
+interface ByCronModelRow {
+  agent_id: string;
+  cron_id: string;
+  cron_name: string | null;
+  model: string;
+  phase: string | null;
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd: number | null;
+}
+
+interface ByPhaseRow {
+  phase: string;
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd: number | null;
+}
+
+interface BySkillRow {
+  kind: string;
+  name: string;
+  runs: bigint;
+  invocations: bigint | null;
+  turns: bigint | null;
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  avg_invoke_context_delta: number | null;
+}
+
+interface BaselineRow {
+  context_fingerprint: string | null;
+  baseline_model: string | null;
+  phase: string | null;
+  runs: bigint;
+  avg_context_tokens: number;
+  min_context_tokens: number;
+  max_context_tokens: number;
+  avg_turns: number | null;
+  avg_tool_calls: number | null;
+  first_seen: Date;
+  last_seen: Date;
+}
+
+// ─── Helpers ─────────────────────────────────────────────────────────────────
+
+function num(v: bigint | null | undefined): number {
+  if (v === null || v === undefined) return 0;
+  return Number(v);
+}
+
+function toAggregate(row: {
+  input: bigint | null;
+  output: bigint | null;
+  cache_read: bigint | null;
+  cache_creation: bigint | null;
+  cost_usd?: number | null;
+}): TokenAggregate {
+  const input = num(row.input);
+  const output = num(row.output);
+  const cacheRead = num(row.cache_read);
+  const cacheCreation = num(row.cache_creation);
+  const result: TokenAggregate = {
+    input,
+    output,
+    cacheRead,
+    cacheCreation,
+    total: input + output + cacheRead + cacheCreation,
+  };
+  if (row.cost_usd !== null && row.cost_usd !== undefined) {
+    result.costUsd = row.cost_usd;
+  }
+  return result;
+}
+
+/**
+ * Build a composable date-range filter fragment for $queryRaw.
+ *
+ * @param from  Lower bound (inclusive).
+ * @param to    Upper bound (exclusive).
+ * @param alias Optional table alias to qualify "startedAt" (e.g. "r" → r."startedAt").
+ *              Pass this whenever the query uses a JOIN so the column reference is
+ *              unambiguous even if the joined table later gains a startedAt column.
+ */
+function dateFilter(
+  from: Date | null,
+  to: Date | null,
+  alias?: string,
+): Prisma.Sql {
+  const col = alias
+    ? Prisma.sql`${Prisma.raw(`${alias}."startedAt"`)}`
+    : Prisma.sql`"startedAt"`;
+  if (from !== null && to !== null) {
+    return Prisma.sql`AND ${col} >= ${from} AND ${col} < ${to}`;
+  }
+  if (from !== null) {
+    return Prisma.sql`AND ${col} >= ${from}`;
+  }
+  if (to !== null) {
+    return Prisma.sql`AND ${col} < ${to}`;
+  }
+  return Prisma.empty;
+}
+
+// ─── Service ─────────────────────────────────────────────────────────────────
+
+export class AgentCronRunStatsService {
+  constructor(private readonly prisma: PrismaClient) {}
+
+  /**
+   * Aggregate cron-run token stats into five dimensions.
+   *
+   * @param from  ISO string — if provided, only runs with startedAt >= from
+   * @param to    ISO string — if provided, only runs with startedAt <  to
+   * @param agentId if provided, only that agent's runs (every dimension);
+   *                omitted preserves the fleet-wide behavior
+   */
+  async query(
+    from?: string,
+    to?: string,
+    agentId?: string,
+  ): Promise<CronRunTokenStats> {
+    const fromDate = from ? new Date(from) : null;
+    const toDate = to ? new Date(to) : null;
+
+    // Every query now aliases AgentCronRun as "r" (all six LEFT/INNER JOIN the
+    // breakdown table), so all filters are qualified against that alias to keep
+    // "startedAt" unambiguous.
+    // The agentId fragment is parameterized and appended to the shared filter,
+    // so it reaches every aggregation block.
+    const filterR = Prisma.sql`${dateFilter(fromDate, toDate, "r")} ${
+      agentId !== undefined
+        ? Prisma.sql`AND r."agentId" = ${agentId}`
+        : Prisma.empty
+    }`;
+
+    const [
+      totalsRows,
+      byAgentRows,
+      byCronRows,
+      byModelRows,
+      dailyRows,
+      byCronModelRows,
+      byPhaseRows,
+      bySkillRows,
+      baselineRows,
+    ] = await Promise.all([
+      this.queryTotals(filterR),
+      this.queryByAgent(filterR),
+      this.queryByCron(filterR),
+      this.queryByModel(filterR),
+      this.queryDaily(filterR),
+      this.queryByCronModel(filterR),
+      this.queryByPhase(filterR),
+      this.queryBySkill(filterR),
+      this.queryBaselines(filterR),
+    ]);
+
+    const totalsRow = totalsRows[0];
+    const totals: TokenAggregate = totalsRow
+      ? toAggregate(totalsRow)
+      : { input: 0, output: 0, cacheRead: 0, cacheCreation: 0, total: 0 };
+
+    const byAgent: KeyedTokenAggregate[] = byAgentRows.map((row) => ({
+      ...toAggregate(row),
+      key: row.agent_id,
+    }));
+
+    const byCron: DoubleKeyedTokenAggregate[] = byCronRows.map((row) => ({
+      ...toAggregate(row),
+      key1: row.agent_id,
+      key2: row.cron_name ?? row.cron_id,
+      phase: row.phase,
+    }));
+
+    const byModel: DoubleKeyedTokenAggregate[] = byModelRows.map((row) => ({
+      ...toAggregate(row),
+      key1: row.agent_id,
+      key2: row.model,
+    }));
+
+    const daily: DailyTokenAggregate[] = dailyRows.map((row) => ({
+      ...toAggregate(row),
+      period: row.period,
+    }));
+
+    const byCronModel: DoubleKeyedTokenAggregate[] = byCronModelRows.map(
+      (row) => ({
+        ...toAggregate(row),
+        key1: `${row.agent_id}:${row.cron_name ?? row.cron_id}`,
+        key2: row.model,
+        phase: row.phase,
+      }),
+    );
+
+    const byPhase: KeyedTokenAggregate[] = byPhaseRows.map((row) => ({
+      ...toAggregate(row),
+      key: row.phase,
+    }));
+
+    const bySkill: SkillUsageAggregate[] = bySkillRows.map((row) => ({
+      kind: row.kind,
+      name: row.name,
+      runs: num(row.runs),
+      invocations: num(row.invocations),
+      turns: num(row.turns),
+      input: num(row.input),
+      output: num(row.output),
+      cacheRead: num(row.cache_read),
+      cacheCreation: num(row.cache_creation),
+      avgInvokeContextDelta: row.avg_invoke_context_delta,
+    }));
+
+    const baselines: ContextBaselineAggregate[] = baselineRows.map((row) => ({
+      contextFingerprint: row.context_fingerprint,
+      baselineModel: row.baseline_model,
+      phase: row.phase,
+      runs: num(row.runs),
+      avgContextTokens: row.avg_context_tokens,
+      minContextTokens: row.min_context_tokens,
+      maxContextTokens: row.max_context_tokens,
+      avgTurns: row.avg_turns,
+      avgToolCalls: row.avg_tool_calls,
+      firstSeen: row.first_seen.toISOString(),
+      lastSeen: row.last_seen.toISOString(),
+    }));
+
+    return {
+      totals,
+      byAgent,
+      byCron,
+      byModel,
+      daily,
+      byCronModel,
+      byPhase,
+      bySkill,
+      baselines,
+    };
+  }
+
+  /**
+   * Per-(phase, contextFingerprint) outcome series for the prompt-audit
+   * before/after comparison. Unlike query(), skipped runs are counted here
+   * (in `skipped` and `skipReasons`) but excluded from the duration, turn,
+   * tool-call, and context-token averages.
+   *
+   * @param agentId if provided, only that agent's runs (both aggregation
+   *                blocks); omitted preserves the fleet-wide behavior
+   */
+  async outcomes(
+    from?: string,
+    to?: string,
+    agentId?: string,
+  ): Promise<CronRunOutcomes> {
+    // Parameterized agentId fragment appended to the shared filter, so it
+    // reaches both queries below.
+    const filter = Prisma.sql`${dateFilter(
+      from ? new Date(from) : null,
+      to ? new Date(to) : null,
+      "r",
+    )} ${
+      agentId !== undefined
+        ? Prisma.sql`AND r."agentId" = ${agentId}`
+        : Prisma.empty
+    }`;
+    const [rows, reasonRows] = await Promise.all([
+      this.prisma.$queryRaw<OutcomeRow[]>`
+        SELECT
+          REGEXP_REPLACE(p.name, '^shipwright-', '') AS phase,
+          r."contextFingerprint"                      AS context_fingerprint,
+          COUNT(*)                                    AS runs,
+          COUNT(*) FILTER (WHERE r.skipped = false AND r.outcome = 'completed') AS completed,
+          COUNT(*) FILTER (WHERE r.skipped = false AND r.outcome = 'failed')    AS failed,
+          COUNT(*) FILTER (WHERE r.skipped = true)                              AS skipped,
+          (AVG(EXTRACT(EPOCH FROM (r."completedAt" - r."startedAt")) * 1000)
+            FILTER (WHERE r.skipped = false AND r."completedAt" IS NOT NULL))::float8 AS avg_duration_ms,
+          (PERCENTILE_CONT(0.5) WITHIN GROUP (
+            ORDER BY EXTRACT(EPOCH FROM (r."completedAt" - r."startedAt")) * 1000
+          ) FILTER (WHERE r.skipped = false AND r."completedAt" IS NOT NULL))::float8 AS p50_duration_ms,
+          (AVG(r.turns) FILTER (WHERE r.skipped = false))::float8                AS avg_turns,
+          (AVG(r."toolCalls") FILTER (WHERE r.skipped = false))::float8          AS avg_tool_calls,
+          (AVG(r."baselineContextTokens") FILTER (WHERE r.skipped = false))::float8 AS avg_context_tokens
+        FROM "AgentCronRun" r
+        LEFT JOIN "AgentCronJob" p ON p.id = r."phaseId"
+        WHERE true
+        ${filter}
+        GROUP BY p.name, r."contextFingerprint"
+        ORDER BY p.name NULLS LAST, r."contextFingerprint" NULLS LAST
+      `,
+      this.prisma.$queryRaw<SkipReasonRow[]>`
+        SELECT
+          REGEXP_REPLACE(p.name, '^shipwright-', '') AS phase,
+          r."contextFingerprint"                      AS context_fingerprint,
+          r."skipReason"                              AS skip_reason,
+          COUNT(*)                                    AS n
+        FROM "AgentCronRun" r
+        LEFT JOIN "AgentCronJob" p ON p.id = r."phaseId"
+        WHERE r.skipped = true
+        ${filter}
+        GROUP BY p.name, r."contextFingerprint", r."skipReason"
+      `,
+    ]);
+
+    const keyOf = (phase: string | null, fp: string | null) =>
+      JSON.stringify([phase, fp]);
+    const reasons = new Map<string, Record<string, number>>();
+    for (const r of reasonRows) {
+      const k = keyOf(r.phase, r.context_fingerprint);
+      const hist = reasons.get(k) ?? {};
+      const reason = r.skip_reason ?? "unknown";
+      hist[reason] = (hist[reason] ?? 0) + Number(r.n);
+      reasons.set(k, hist);
+    }
+
+    return {
+      series: rows.map((row) => ({
+        phase: row.phase,
+        contextFingerprint: row.context_fingerprint,
+        runs: Number(row.runs),
+        completed: Number(row.completed),
+        failed: Number(row.failed),
+        skipped: Number(row.skipped),
+        skipReasons:
+          reasons.get(keyOf(row.phase, row.context_fingerprint)) ?? {},
+        avgDurationMs: row.avg_duration_ms,
+        p50DurationMs: row.p50_duration_ms,
+        avgTurns: row.avg_turns,
+        avgToolCalls: row.avg_tool_calls,
+        avgContextTokens: row.avg_context_tokens,
+      })),
+    };
+  }
+
+  // ─── Private query methods ──────────────────────────────────────────────────
+
+  private queryTotals(filter: Prisma.Sql): Promise<TotalsRow[]> {
+    // All token/cost data lives in AgentCronRunModelBreakdown. A LEFT JOIN keeps
+    // runs without breakdown rows in scope for the skipped/date filters (they
+    // simply contribute NULL, i.e. zero, to every SUM).
+    return this.prisma.$queryRaw<TotalsRow[]>`
+      SELECT
+        SUM(b."inputTokens")         AS input,
+        SUM(b."outputTokens")        AS output,
+        SUM(b."cacheReadTokens")     AS cache_read,
+        SUM(b."cacheCreationTokens") AS cache_creation,
+        SUM(b."costUsd")             AS cost_usd
+      FROM "AgentCronRun" r
+      LEFT JOIN "AgentCronRunModelBreakdown" b ON b."cronRunId" = r.id
+      WHERE r.skipped = false
+      ${filter}
+    `;
+  }
+
+  private queryByAgent(filter: Prisma.Sql): Promise<ByAgentRow[]> {
+    return this.prisma.$queryRaw<ByAgentRow[]>`
+      SELECT
+        r."agentId"                  AS agent_id,
+        SUM(b."inputTokens")         AS input,
+        SUM(b."outputTokens")        AS output,
+        SUM(b."cacheReadTokens")     AS cache_read,
+        SUM(b."cacheCreationTokens") AS cache_creation,
+        SUM(b."costUsd")             AS cost_usd
+      FROM "AgentCronRun" r
+      LEFT JOIN "AgentCronRunModelBreakdown" b ON b."cronRunId" = r.id
+      WHERE r.skipped = false
+      ${filter}
+      GROUP BY r."agentId"
+      ORDER BY r."agentId"
+    `;
+  }
+
+  private queryByCron(filter: Prisma.Sql): Promise<ByCronRow[]> {
+    // GROUP BY includes p.name (LPC-3.1, via phaseId): Postgres treats NULL
+    // as a single group, so legacy/unreconciled runs (phaseId IS NULL) for a
+    // cron still collapse into one row — identical to the pre-WL-3.5
+    // cronId-only grouping. Runs that do carry a phaseId produce one row per
+    // (cronId, phase). The "shipwright-" prefix is stripped from p.name so
+    // the label matches exactly what the old r.phase string column used to
+    // contain (e.g. "dev-task", not "shipwright-dev-task").
+    return this.prisma.$queryRaw<ByCronRow[]>`
+      SELECT
+        r."agentId"                              AS agent_id,
+        r."cronId"                               AS cron_id,
+        j.name                                   AS cron_name,
+        REGEXP_REPLACE(p.name, '^shipwright-', '') AS phase,
+        SUM(b."inputTokens")         AS input,
+        SUM(b."outputTokens")        AS output,
+        SUM(b."cacheReadTokens")     AS cache_read,
+        SUM(b."cacheCreationTokens") AS cache_creation,
+        SUM(b."costUsd")             AS cost_usd
+      FROM "AgentCronRun" r
+      LEFT JOIN "AgentCronJob" j ON j.id = r."cronId"
+      LEFT JOIN "AgentCronJob" p ON p.id = r."phaseId"
+      LEFT JOIN "AgentCronRunModelBreakdown" b ON b."cronRunId" = r.id
+      WHERE r.skipped = false
+      ${filter}
+      GROUP BY r."agentId", r."cronId", j.name, p.name
+      ORDER BY r."agentId", r."cronId", p.name
+    `;
+  }
+
+  private queryByModel(filter: Prisma.Sql): Promise<ByModelRow[]> {
+    // All per-model data comes from AgentCronRunModelBreakdown.
+    // Runs without breakdown rows simply have no byModel data.
+    return this.prisma.$queryRaw<ByModelRow[]>`
+      SELECT
+        r."agentId"                AS agent_id,
+        b.model                    AS model,
+        SUM(b."inputTokens")       AS input,
+        SUM(b."outputTokens")      AS output,
+        SUM(b."cacheReadTokens")   AS cache_read,
+        SUM(b."cacheCreationTokens") AS cache_creation,
+        SUM(b."costUsd")           AS cost_usd
+      FROM "AgentCronRun" r
+      INNER JOIN "AgentCronRunModelBreakdown" b ON b."cronRunId" = r.id
+      WHERE r.skipped = false
+      ${filter}
+      GROUP BY r."agentId", b.model
+      ORDER BY r."agentId", b.model
+    `;
+  }
+
+  private queryDaily(filter: Prisma.Sql): Promise<DailyRow[]> {
+    return this.prisma.$queryRaw<DailyRow[]>`
+      SELECT
+        TO_CHAR(DATE(r."startedAt"), 'YYYY-MM-DD') AS period,
+        SUM(b."inputTokens")         AS input,
+        SUM(b."outputTokens")        AS output,
+        SUM(b."cacheReadTokens")     AS cache_read,
+        SUM(b."cacheCreationTokens") AS cache_creation,
+        SUM(b."costUsd")             AS cost_usd
+      FROM "AgentCronRun" r
+      LEFT JOIN "AgentCronRunModelBreakdown" b ON b."cronRunId" = r.id
+      WHERE r.skipped = false
+      ${filter}
+      GROUP BY DATE(r."startedAt")
+      ORDER BY DATE(r."startedAt")
+    `;
+  }
+
+  private queryByCronModel(filter: Prisma.Sql): Promise<ByCronModelRow[]> {
+    // GROUP BY includes p.name (LPC-3.1, via phaseId) — same
+    // NULL-collapses-to-one-group reasoning as queryByCron above.
+    return this.prisma.$queryRaw<ByCronModelRow[]>`
+      SELECT
+        r."agentId"                              AS agent_id,
+        r."cronId"                               AS cron_id,
+        j.name                                   AS cron_name,
+        b.model                                  AS model,
+        REGEXP_REPLACE(p.name, '^shipwright-', '') AS phase,
+        SUM(b."inputTokens")           AS input,
+        SUM(b."outputTokens")          AS output,
+        SUM(b."cacheReadTokens")       AS cache_read,
+        SUM(b."cacheCreationTokens")   AS cache_creation,
+        SUM(b."costUsd")               AS cost_usd
+      FROM "AgentCronRun" r
+      LEFT JOIN "AgentCronJob" j ON j.id = r."cronId"
+      LEFT JOIN "AgentCronJob" p ON p.id = r."phaseId"
+      INNER JOIN "AgentCronRunModelBreakdown" b ON b."cronRunId" = r.id
+      WHERE r.skipped = false
+      ${filter}
+      GROUP BY r."agentId", r."cronId", j.name, b.model, p.name
+      ORDER BY r."agentId", r."cronId", b.model, p.name
+    `;
+  }
+
+  private queryByPhase(filter: Prisma.Sql): Promise<ByPhaseRow[]> {
+    // Runs with no phaseId (legacy five-job crons, or an unreconciled agent)
+    // are excluded — phase is additive grouping, not a replacement for
+    // totals/byAgent/etc. Label derived by LEFT JOINing "AgentCronJob" (alias
+    // p) on r."phaseId" = p.id and stripping the "shipwright-" prefix from
+    // p.name, matching what r.phase used to contain directly.
+    return this.prisma.$queryRaw<ByPhaseRow[]>`
+      SELECT
+        REGEXP_REPLACE(p.name, '^shipwright-', '') AS phase,
+        SUM(b."inputTokens")         AS input,
+        SUM(b."outputTokens")        AS output,
+        SUM(b."cacheReadTokens")     AS cache_read,
+        SUM(b."cacheCreationTokens") AS cache_creation,
+        SUM(b."costUsd")             AS cost_usd
+      FROM "AgentCronRun" r
+      LEFT JOIN "AgentCronJob" p ON p.id = r."phaseId"
+      LEFT JOIN "AgentCronRunModelBreakdown" b ON b."cronRunId" = r.id
+      WHERE r.skipped = false AND r."phaseId" IS NOT NULL
+      ${filter}
+      GROUP BY p.name
+      ORDER BY p.name
+    `;
+  }
+
+  private queryBySkill(filter: Prisma.Sql): Promise<BySkillRow[]> {
+    // Skipped runs excluded (same rule as every token dimension). AVG ignores
+    // NULL invokeContextDelta rows, so a skill never measured yields NULL.
+    return this.prisma.$queryRaw<BySkillRow[]>`
+      SELECT
+        u.kind                              AS kind,
+        u.name                              AS name,
+        COUNT(DISTINCT r.id)                AS runs,
+        SUM(u.invocations)                  AS invocations,
+        SUM(u.turns)                        AS turns,
+        SUM(u."inputTokens")                AS input,
+        SUM(u."outputTokens")               AS output,
+        SUM(u."cacheReadTokens")            AS cache_read,
+        SUM(u."cacheCreationTokens")        AS cache_creation,
+        AVG(u."invokeContextDelta")::float8 AS avg_invoke_context_delta
+      FROM "AgentCronRun" r
+      INNER JOIN "AgentCronRunSkillUsage" u ON u."cronRunId" = r.id
+      WHERE r.skipped = false
+      ${filter}
+      GROUP BY u.kind, u.name
+      ORDER BY u.kind, u.name
+    `;
+  }
+
+  private queryBaselines(filter: Prisma.Sql): Promise<BaselineRow[]> {
+    // Skipped runs are deliberately included; runs with no baseline are not.
+    // NULL fingerprint/model/phase each collapse into a single group.
+    return this.prisma.$queryRaw<BaselineRow[]>`
+      SELECT
+        r."contextFingerprint"                       AS context_fingerprint,
+        r."baselineModel"                            AS baseline_model,
+        REGEXP_REPLACE(p.name, '^shipwright-', '')   AS phase,
+        COUNT(*)                                     AS runs,
+        AVG(r."baselineContextTokens")::float8       AS avg_context_tokens,
+        MIN(r."baselineContextTokens")               AS min_context_tokens,
+        MAX(r."baselineContextTokens")               AS max_context_tokens,
+        AVG(r.turns)::float8                         AS avg_turns,
+        AVG(r."toolCalls")::float8                   AS avg_tool_calls,
+        MIN(r."startedAt")                           AS first_seen,
+        MAX(r."startedAt")                           AS last_seen
+      FROM "AgentCronRun" r
+      LEFT JOIN "AgentCronJob" p ON p.id = r."phaseId"
+      WHERE r."baselineContextTokens" IS NOT NULL
+      ${filter}
+      GROUP BY r."contextFingerprint", r."baselineModel", p.name
+      ORDER BY MIN(r."startedAt"), r."contextFingerprint", r."baselineModel", p.name
+    `;
+  }
+}