@desplega.ai/agent-swarm 1.79.4 → 1.80.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@desplega.ai/agent-swarm",
-  "version": "1.79.4",
+  "version": "1.80.1",
   "description": "Multi-agent orchestration for Claude Code, Codex, Gemini CLI, and other AI coding assistants",
   "license": "MIT",
   "author": "desplega.sh <contact@desplega.sh>",
@@ -42,8 +42,10 @@
   },
   "scripts": {
     "build:pi-skills": "bun run plugin/build-pi-skills.ts",
+    "build:script-types": "bun run scripts/bundle-script-types.ts",
     "tsc:check": "bun tsc --noEmit",
     "check:db-boundary": "bash scripts/check-db-boundary.sh",
+    "check:api-key-boundary": "bash scripts/check-api-key-boundary.sh",
     "cli": "bun src/cli.tsx",
     "hook": "bun src/hooks/hook.ts",
     "claude": "bun src/cli.tsx claude",
@@ -73,6 +75,7 @@
     "deploy:docker": "bun deploy/docker-push.ts",
     "e2e:workflows": "bun scripts/e2e-workflow-test.ts",
     "e2e:workflows:docker": "bun scripts/e2e-workflow-test.ts --with-docker",
+    "e2e:otel:jaeger": "bun scripts/e2e-otel-jaeger.ts",
     "docs:mcp": "bun scripts/generate-mcp-docs.ts",
     "docs:openapi": "bun scripts/generate-openapi.ts",
     "docs:business-use": "bun scripts/generate-business-use-docs.ts",
@@ -104,13 +107,18 @@
     "@desplega.ai/localtunnel": "^2.2.0",
     "@inkjs/ui": "^2.0.0",
     "@linear/sdk": "^77.0.0",
-    "@earendil-works/pi-agent-core": "^0.74.0",
-    "@earendil-works/pi-ai": "^0.74.0",
-    "@earendil-works/pi-coding-agent": "^0.74.0",
+    "@earendil-works/pi-agent-core": "^0.75.3",
+    "@earendil-works/pi-ai": "^0.75.3",
+    "@earendil-works/pi-coding-agent": "^0.75.3",
     "@modelcontextprotocol/sdk": "^1.25.1",
-    "@openai/codex-sdk": "^0.128.0",
-    "@opencode-ai/sdk": "^1.14.30",
+    "@openai/codex-sdk": "^0.130.0",
+    "@opencode-ai/sdk": "^1.15.4",
     "@openfort/openfort-node": "^0.9.1",
+    "@opentelemetry/api": "^1.9.1",
+    "@opentelemetry/exporter-trace-otlp-http": "^0.218.0",
+    "@opentelemetry/resources": "^2.7.1",
+    "@opentelemetry/sdk-node": "^0.218.0",
+    "@opentelemetry/semantic-conventions": "^1.41.1",
     "@slack/bolt": "^4.6.0",
     "@types/react": "^19.2.7",
     "@x402/core": "^2.5.0",

package/src/artifact-sdk/server.ts

@@ -1,5 +1,6 @@
 import { Hono } from "hono";
 import { serveStatic } from "hono/bun";
+import { getApiKey } from "../utils/api-key";
 import { BROWSER_SDK_JS } from "./browser-sdk";
 import { getAvailablePort } from "./port";
 import { createTunnel } from "./tunnel";
@@ -23,7 +24,7 @@ export interface ArtifactServer {
 
 export function createArtifactServer(opts: ArtifactServerOptions): ArtifactServer {
   const agentId = process.env.AGENT_ID || "unknown";
-  const apiKey = process.env.API_KEY || "";
+  const apiKey = getApiKey();
   const mcpBaseUrl = process.env.MCP_BASE_URL || `http://localhost:${process.env.PORT || "3013"}`;
 
   const app = new Hono();

package/src/be/db.ts

@@ -362,7 +362,7 @@ function ensureAgentProfileColumns(database: Database): void {
   }
 }
 
-function computeContentHash(content: string): string {
+export function computeContentHash(content: string): string {
   const hasher = new Bun.CryptoHasher("sha256");
   hasher.update(content);
   return hasher.digest("hex");
@@ -980,7 +980,7 @@ type AgentTaskRow = {
   progress: string | null;
   compactionCount: number | null;
   peakContextPercent: number | null;
-  totalContextTokensUsed: number | null;
+  peakContextTokens: number | null;
   contextWindowSize: number | null;
   was_paused: number;
   credentialKeySuffix: string | null;
@@ -1036,7 +1036,7 @@ function rowToAgentTask(row: AgentTaskRow): AgentTask {
     contextKey: row.contextKey ?? undefined,
     compactionCount: row.compactionCount ?? undefined,
     peakContextPercent: row.peakContextPercent ?? undefined,
-    totalContextTokensUsed: row.totalContextTokensUsed ?? undefined,
+    peakContextTokens: row.peakContextTokens ?? undefined,
     contextWindowSize: row.contextWindowSize ?? undefined,
     createdAt: row.createdAt,
     lastUpdatedAt: row.lastUpdatedAt,
@@ -3761,8 +3761,11 @@ type SessionCostRow = {
   outputTokens: number;
   cacheReadTokens: number;
   cacheWriteTokens: number;
+  // Migration 063 additions:
+  reasoningOutputTokens: number;
+  thinkingTokens: number;
   durationMs: number;
-  numTurns: number;
+  numTurns: number | null;
   model: string;
   isError: number;
   costSource: string;
@@ -3780,6 +3783,8 @@ function rowToSessionCost(row: SessionCostRow): SessionCost {
     outputTokens: row.outputTokens,
     cacheReadTokens: row.cacheReadTokens,
     cacheWriteTokens: row.cacheWriteTokens,
+    reasoningOutputTokens: row.reasoningOutputTokens ?? 0,
+    thinkingTokens: row.thinkingTokens ?? 0,
     durationMs: row.durationMs,
     numTurns: row.numTurns,
     model: row.model,
@@ -3803,15 +3808,24 @@ const sessionCostQueries = {
         number,
         number,
         number,
-        number,
-        number,
-        string,
-        number,
-        string,
+        number, // reasoningOutputTokens
+        number, // thinkingTokens
+        number, // durationMs
+        number | null, // numTurns
+        string, // model
+        number, // isError
+        string, // costSource
       ]
     >(
-      `INSERT INTO session_costs (id, sessionId, taskId, agentId, totalCostUsd, inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, durationMs, numTurns, model, isError, costSource, createdAt)
-       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, strftime('%Y-%m-%dT%H:%M:%fZ', 'now'))`,
+      `INSERT INTO session_costs (
+         id, sessionId, taskId, agentId,
+         totalCostUsd, inputTokens, outputTokens,
+         cacheReadTokens, cacheWriteTokens,
+         reasoningOutputTokens, thinkingTokens,
+         durationMs, numTurns, model, isError,
+         costSource, createdAt
+       )
+       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, strftime('%Y-%m-%dT%H:%M:%fZ', 'now'))`,
     ),
 
   getByTaskId: () =>
@@ -3839,16 +3853,22 @@ export interface CreateSessionCostInput {
   outputTokens?: number;
   cacheReadTokens?: number;
   cacheWriteTokens?: number;
+  // Migration 063 additions — adapters that have these numbers should pass
+  // them; defaulting to 0 preserves the old write shape for callers that don't.
+  reasoningOutputTokens?: number;
+  thinkingTokens?: number;
   durationMs: number;
-  numTurns: number;
+  // Nullable: some adapters (claude when num_turns is absent) can't honestly
+  // report a turn count; we prefer null over a faked 1.
+  numTurns: number | null;
   model: string;
   isError?: boolean;
   /**
-   * Phase 6: where the recorded `totalCostUsd` came from.
+   * Phase 6 (migration 063 added 'unpriced'): where `totalCostUsd` came from.
    *  - 'harness'        — value reported by the harness as-is (default).
-   *  - 'pricing-table'  — value recomputed by the API from `pricing` rows
-   *                       (Codex when DB pricing rows exist for all three
-   *                       token classes).
+   *  - 'pricing-table'  — value recomputed by the API from `pricing` rows.
+   *  - 'unpriced'       — recompute attempted but no matching pricing rows;
+   *                       `totalCostUsd` is whatever the worker submitted.
    */
   costSource?: SessionCostSource;
 }
@@ -3856,6 +3876,8 @@ export interface CreateSessionCostInput {
 export function createSessionCost(input: CreateSessionCostInput): SessionCost {
   const id = crypto.randomUUID();
   const costSource: SessionCostSource = input.costSource ?? "harness";
+  const reasoningOutputTokens = input.reasoningOutputTokens ?? 0;
+  const thinkingTokens = input.thinkingTokens ?? 0;
   sessionCostQueries
     .insert()
     .run(
@@ -3868,6 +3890,8 @@ export function createSessionCost(input: CreateSessionCostInput): SessionCost {
       input.outputTokens ?? 0,
       input.cacheReadTokens ?? 0,
       input.cacheWriteTokens ?? 0,
+      reasoningOutputTokens,
+      thinkingTokens,
       input.durationMs,
       input.numTurns,
       input.model,
@@ -3885,6 +3909,8 @@ export function createSessionCost(input: CreateSessionCostInput): SessionCost {
     outputTokens: input.outputTokens ?? 0,
     cacheReadTokens: input.cacheReadTokens ?? 0,
     cacheWriteTokens: input.cacheWriteTokens ?? 0,
+    reasoningOutputTokens,
+    thinkingTokens,
     durationMs: input.durationMs,
     numTurns: input.numTurns,
     model: input.model,
@@ -4110,16 +4136,33 @@ export interface DashboardCostSummary {
 }
 
 export function getDashboardCostSummary(): DashboardCostSummary {
+  // Phase 13: compute the date boundaries in TS and pass them as ISO 8601
+  // strings. `session_costs.createdAt` is a TEXT ISO 8601 column; lexicographic
+  // comparison on ISO 8601 sorts correctly, so the comparison works as long
+  // as both sides are the same shape. The old code compared an ISO string
+  // (`2026-05-15T03:45:12.123Z`) against `date('now')` (which returns the
+  // string `2026-05-15`) — lexicographically `2026-05-15T...` > `2026-05-15`,
+  // so post-midnight rows correctly counted, BUT rows whose ISO began with
+  // the EXACT bare-date string would fail the `>=` check inconsistently
+  // depending on millisecond precision. Use a proper ISO-millisecond boundary
+  // for both halves so the comparison is unambiguous.
+  const now = new Date();
+  const startOfDayUtc = new Date(
+    Date.UTC(now.getUTCFullYear(), now.getUTCMonth(), now.getUTCDate()),
+  ).toISOString();
+  const startOfMonthUtc = new Date(
+    Date.UTC(now.getUTCFullYear(), now.getUTCMonth(), 1),
+  ).toISOString();
   type CostRow = { costToday: number; costMtd: number };
   const row = getDb()
-    .prepare<CostRow, []>(
+    .prepare<CostRow, [string, string]>(
       `SELECT
-        COALESCE(SUM(CASE WHEN createdAt >= date('now') THEN totalCostUsd ELSE 0 END), 0) as costToday,
+        COALESCE(SUM(CASE WHEN createdAt >= ? THEN totalCostUsd ELSE 0 END), 0) as costToday,
         COALESCE(SUM(totalCostUsd), 0) as costMtd
       FROM session_costs
-      WHERE createdAt >= date('now', 'start of month')`,
+      WHERE createdAt >= ?`,
     )
-    .get();
+    .get(startOfDayUtc, startOfMonthUtc);
 
   return row ?? { costToday: 0, costMtd: 0 };
 }
@@ -8245,6 +8288,8 @@ type ContextSnapshotRow = {
   preCompactTokens: number | null;
   cumulativeInputTokens: number;
   cumulativeOutputTokens: number;
+  // Migration 063 — see ContextFormulaSchema in src/types.ts for the value set.
+  contextFormula: string | null;
   createdAt: string;
 };
 
@@ -8258,10 +8303,11 @@ function rowToContextSnapshot(row: ContextSnapshotRow): ContextSnapshot {
     contextTotalTokens: row.contextTotalTokens ?? undefined,
     contextPercent: row.contextPercent ?? undefined,
     eventType: row.eventType,
-    compactTrigger: (row.compactTrigger as "auto" | "manual" | null) ?? undefined,
+    compactTrigger: (row.compactTrigger as "auto" | "manual" | "auto-inferred" | null) ?? undefined,
     preCompactTokens: row.preCompactTokens ?? undefined,
     cumulativeInputTokens: row.cumulativeInputTokens,
     cumulativeOutputTokens: row.cumulativeOutputTokens,
+    contextFormula: (row.contextFormula as ContextSnapshot["contextFormula"]) ?? undefined,
     createdAt: row.createdAt,
   };
 }
@@ -8283,11 +8329,12 @@ const contextSnapshotQueries = {
         number | null,
         number,
         number,
+        string | null, // contextFormula (migration 063)
         string,
       ]
     >(
-      `INSERT INTO task_context_snapshots (id, taskId, agentId, sessionId, contextUsedTokens, contextTotalTokens, contextPercent, eventType, compactTrigger, preCompactTokens, cumulativeInputTokens, cumulativeOutputTokens, createdAt)
-       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,
+      `INSERT INTO task_context_snapshots (id, taskId, agentId, sessionId, contextUsedTokens, contextTotalTokens, contextPercent, eventType, compactTrigger, preCompactTokens, cumulativeInputTokens, cumulativeOutputTokens, contextFormula, createdAt)
+       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,
     ),
 
   getByTaskId: () =>
@@ -8309,10 +8356,12 @@ export interface CreateContextSnapshotInput {
   contextTotalTokens?: number;
   contextPercent?: number;
   eventType: ContextSnapshotEventType;
-  compactTrigger?: "auto" | "manual";
+  compactTrigger?: "auto" | "manual" | "auto-inferred";
   preCompactTokens?: number;
   cumulativeInputTokens?: number;
   cumulativeOutputTokens?: number;
+  // Migration 063 — adapter-supplied formula tag.
+  contextFormula?: ContextSnapshot["contextFormula"];
 }
 
 export function createContextSnapshot(input: CreateContextSnapshotInput): ContextSnapshot {
@@ -8334,6 +8383,7 @@ export function createContextSnapshot(input: CreateContextSnapshotInput): Contex
       input.preCompactTokens ?? null,
       input.cumulativeInputTokens ?? 0,
       input.cumulativeOutputTokens ?? 0,
+      input.contextFormula ?? null,
       now,
     );
 
@@ -8347,10 +8397,15 @@ export function createContextSnapshot(input: CreateContextSnapshotInput): Contex
       .run(input.contextPercent, input.taskId);
   }
 
-  // Keep totalContextTokensUsed up to date with the latest known value
+  // Migration 063: peakContextTokens is monotonic-max across snapshots, not a
+  // rolling latest. Mirrors Claude Code's status-line "peak context" semantic.
   if (input.contextUsedTokens != null) {
     getDb()
-      .prepare("UPDATE agent_tasks SET totalContextTokensUsed = ? WHERE id = ?")
+      .prepare(
+        `UPDATE agent_tasks
+         SET peakContextTokens = MAX(COALESCE(peakContextTokens, 0), ?)
+         WHERE id = ?`,
+      )
       .run(input.contextUsedTokens, input.taskId);
   }
 
@@ -8362,9 +8417,17 @@ export function createContextSnapshot(input: CreateContextSnapshotInput): Contex
       .run(input.taskId);
   }
 
-  if (input.eventType === "completion" && input.contextTotalTokens != null) {
+  // Phase 10: set contextWindowSize on the FIRST snapshot that carries one
+  // (was previously gated on eventType === 'completion', meaning the UI saw
+  // NULL throughout running tasks). Subsequent snapshots leave it alone — the
+  // window doesn't change mid-session.
+  if (input.contextTotalTokens != null) {
     getDb()
-      .prepare("UPDATE agent_tasks SET contextWindowSize = ? WHERE id = ?")
+      .prepare(
+        `UPDATE agent_tasks
+         SET contextWindowSize = ?
+         WHERE id = ? AND contextWindowSize IS NULL`,
+      )
       .run(input.contextTotalTokens, input.taskId);
   }
 
@@ -8381,6 +8444,7 @@ export function createContextSnapshot(input: CreateContextSnapshotInput): Contex
     preCompactTokens: input.preCompactTokens,
     cumulativeInputTokens: input.cumulativeInputTokens ?? 0,
     cumulativeOutputTokens: input.cumulativeOutputTokens ?? 0,
+    contextFormula: input.contextFormula,
     createdAt: now,
   };
 }
@@ -8396,7 +8460,8 @@ export function getContextSnapshotsBySessionId(sessionId: string, limit = 500):
 export interface ContextSummary {
   compactionCount: number;
   peakContextPercent: number | null;
-  totalContextTokensUsed: number | null;
+  // Migration 063: renamed from totalContextTokensUsed.
+  peakContextTokens: number | null;
   contextWindowSize: number | null;
   snapshotCount: number;
 }
@@ -8412,7 +8477,7 @@ export function getContextSummaryByTaskId(taskId: string): ContextSummary {
   return {
     compactionCount: task?.compactionCount ?? 0,
     peakContextPercent: task?.peakContextPercent ?? null,
-    totalContextTokensUsed: task?.totalContextTokensUsed ?? null,
+    peakContextTokens: task?.peakContextTokens ?? null,
     contextWindowSize: task?.contextWindowSize ?? null,
     snapshotCount: countRow?.cnt ?? 0,
   };
@@ -8635,6 +8700,12 @@ export function getKeyCostSummary(keyType?: string): KeyCostSummary[] {
   }
 
   const where = conditions.length > 0 ? `WHERE ${conditions.join(" AND ")}` : "";
+  // Phase 13: INNER JOIN -> LEFT JOIN. The `WHERE t.credentialKeySuffix IS NOT NULL`
+  // still filters out rows whose taskId doesn't link to a task with credentials,
+  // but switching to LEFT JOIN means a future change that drops the WHERE
+  // (or a debugging query that wants orphan rows visible) doesn't silently
+  // disappear them. Equivalent for the current `WHERE … IS NOT NULL` filter;
+  // makes the query's intent (cost rows owned by a credential) explicit.
   return db
     .prepare<KeyCostSummary, string[]>(
       `SELECT
@@ -8645,7 +8716,7 @@ export function getKeyCostSummary(keyType?: string): KeyCostSummary[] {
         COALESCE(SUM(sc.outputTokens), 0) as totalOutputTokens,
         COUNT(DISTINCT sc.taskId) as taskCount
       FROM session_costs sc
-      JOIN agent_tasks t ON sc.taskId = t.id
+      LEFT JOIN agent_tasks t ON sc.taskId = t.id
       ${where}
       GROUP BY t.credentialKeyType, t.credentialKeySuffix`,
     )

package/src/be/migrations/063_cost_context_schema_relax.sql

@@ -0,0 +1,133 @@
+-- 063_cost_context_schema_relax.sql
+-- Phase 1 of the context & cost tracking fixes plan (2026-05-15).
+--
+-- This migration unblocks every downstream phase by:
+--   * Dropping the brittle CHECK constraints on `pricing.provider` and
+--     `pricing.token_class` so we can seed rows for all 7 providers
+--     (claude, claude-managed, codex, pi, opencode, devin, gemini) and the
+--     extra token classes (`cache_write`, `runtime_hour`, `acu`). Zod
+--     validation at the application boundary (`PricingProviderSchema`,
+--     `PricingTokenClassSchema` in `src/types.ts`) keeps the actual safety
+--     guarantee — the CHECKs added drift risk for no real benefit.
+--   * Renaming the misleading `agent_tasks.totalContextTokensUsed` column
+--     to `peakContextTokens` to match its new monotonic-max semantic
+--     (mirrors Claude Code's status-line "peak context" idea).
+--   * Recording the `contextFormula` used by the adapter that emitted a
+--     given snapshot so we can tell apples from oranges across providers.
+--   * Adding `reasoningOutputTokens` (codex reasoning models) and
+--     `thinkingTokens` (claude extended thinking) columns to `session_costs`
+--     so we stop dropping those numbers on the floor.
+--
+-- SQLite CHECK constraints can't be modified in place, so the `pricing` and
+-- `task_context_snapshots` shape changes use the standard
+-- create-new / copy / drop / rename dance. Existing rows are preserved.
+--
+-- Forward-only — no down migration. If you need to revert, write a new
+-- migration that walks the schema forward to the desired state.
+
+-- ---------------------------------------------------------------------------
+-- 1. Relax `pricing` CHECK constraints (drop them entirely; Zod validates).
+-- ---------------------------------------------------------------------------
+
+CREATE TABLE pricing_new (
+  provider TEXT NOT NULL,
+  model TEXT NOT NULL,
+  token_class TEXT NOT NULL,
+  effective_from INTEGER NOT NULL,
+  price_per_million_usd REAL NOT NULL,
+  createdAt INTEGER NOT NULL,
+  lastUpdatedAt INTEGER NOT NULL,
+  PRIMARY KEY (provider, model, token_class, effective_from)
+);
+
+INSERT INTO pricing_new (provider, model, token_class, effective_from, price_per_million_usd, createdAt, lastUpdatedAt)
+SELECT provider, model, token_class, effective_from, price_per_million_usd, createdAt, lastUpdatedAt
+FROM pricing;
+
+DROP TABLE pricing;
+ALTER TABLE pricing_new RENAME TO pricing;
+
+-- Re-create the index the original `pricing` table had (matches 046:54-55).
+CREATE INDEX IF NOT EXISTS idx_pricing_lookup
+  ON pricing (provider, model, token_class, effective_from DESC);
+
+-- ---------------------------------------------------------------------------
+-- 2. Rename agent_tasks.totalContextTokensUsed -> peakContextTokens.
+--    SQLite >= 3.25 supports RENAME COLUMN; bun:sqlite is well past that.
+-- ---------------------------------------------------------------------------
+
+ALTER TABLE agent_tasks RENAME COLUMN totalContextTokensUsed TO peakContextTokens;
+
+-- ---------------------------------------------------------------------------
+-- 3. Add contextFormula column to task_context_snapshots.
+--    Using a plain TEXT column (no CHECK) so the adapter side can add new
+--    formulas without an accompanying migration; Zod enum validates writes.
+--    Values today:
+--      'input-cache-output'    — unified formula (post-Phase 9)
+--      'input-cache-no-output' — pre-unification claude formula
+--      'input-output-no-cache' — pre-unification claude-managed formula
+--      'peak-proxy'            — pre-unification codex formula
+--      'pi-delegated'          — context numbers come from the pi-ai SDK
+--      'harness-reported'      — context numbers come from a harness API (devin)
+--      'unknown'               — pre-migration backfill or adapter didn't tag
+-- ---------------------------------------------------------------------------
+
+ALTER TABLE task_context_snapshots ADD COLUMN contextFormula TEXT;
+UPDATE task_context_snapshots SET contextFormula = 'unknown' WHERE contextFormula IS NULL;
+
+-- ---------------------------------------------------------------------------
+-- 4. Rewrite session_costs to:
+--    a) drop the costSource CHECK (we need 'unpriced' as a third value);
+--    b) add reasoningOutputTokens + thinkingTokens columns we previously
+--       dropped on the floor.
+--    SQLite can't relax a CHECK in-place — table-rewrite dance, same pattern
+--    as the pricing table above. FKs and indexes are restored after rename.
+-- ---------------------------------------------------------------------------
+
+CREATE TABLE session_costs_new (
+    id TEXT PRIMARY KEY,
+    sessionId TEXT NOT NULL,
+    taskId TEXT,
+    agentId TEXT NOT NULL,
+    totalCostUsd REAL NOT NULL,
+    inputTokens INTEGER NOT NULL DEFAULT 0,
+    outputTokens INTEGER NOT NULL DEFAULT 0,
+    cacheReadTokens INTEGER NOT NULL DEFAULT 0,
+    -- Migration 063: nullable. Codex SDK can't surface cache writes, so we
+    -- store null instead of faking a 0 that mixes with real zeros.
+    cacheWriteTokens INTEGER DEFAULT 0,
+    durationMs INTEGER NOT NULL,
+    -- Migration 063: nullable. Claude when `num_turns` is absent can't honestly
+    -- report a turn count; null is preferred over a faked 1.
+    numTurns INTEGER,
+    model TEXT NOT NULL,
+    isError INTEGER NOT NULL DEFAULT 0,
+    costSource TEXT NOT NULL DEFAULT 'harness',
+    reasoningOutputTokens INTEGER NOT NULL DEFAULT 0,
+    thinkingTokens INTEGER NOT NULL DEFAULT 0,
+    createdAt TEXT NOT NULL,
+    FOREIGN KEY (agentId) REFERENCES agents(id) ON DELETE CASCADE,
+    FOREIGN KEY (taskId) REFERENCES agent_tasks(id) ON DELETE SET NULL
+);
+
+INSERT INTO session_costs_new (
+    id, sessionId, taskId, agentId, totalCostUsd,
+    inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens,
+    durationMs, numTurns, model, isError, costSource,
+    reasoningOutputTokens, thinkingTokens, createdAt
+)
+SELECT
+    id, sessionId, taskId, agentId, totalCostUsd,
+    inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens,
+    durationMs, numTurns, model, isError, costSource,
+    0, 0, createdAt
+FROM session_costs;
+
+DROP TABLE session_costs;
+ALTER TABLE session_costs_new RENAME TO session_costs;
+
+-- Recreate indexes (mirrors 001_initial.sql:360-363).
+CREATE INDEX IF NOT EXISTS idx_session_costs_createdAt ON session_costs(createdAt);
+CREATE INDEX IF NOT EXISTS idx_session_costs_taskId ON session_costs(taskId);
+CREATE INDEX IF NOT EXISTS idx_session_costs_agentId ON session_costs(agentId);
+CREATE INDEX IF NOT EXISTS idx_session_costs_agent_createdAt ON session_costs(agentId, createdAt);

package/src/be/migrations/064_scripts.sql

@@ -0,0 +1,39 @@
+CREATE TABLE scripts (
+  id TEXT PRIMARY KEY,
+  name TEXT NOT NULL,
+  scope TEXT NOT NULL CHECK(scope IN ('global', 'agent')),
+  scopeId TEXT,
+  source TEXT NOT NULL,
+  description TEXT NOT NULL,
+  intent TEXT NOT NULL,
+  signatureJson TEXT NOT NULL,
+  contentHash TEXT NOT NULL,
+  version INTEGER NOT NULL DEFAULT 1,
+  isScratch INTEGER NOT NULL DEFAULT 0,
+  typeChecked INTEGER NOT NULL DEFAULT 0,
+  fsMode TEXT NOT NULL DEFAULT 'none' CHECK(fsMode IN ('none', 'workspace-rw')),
+  createdByAgentId TEXT,
+  createdAt TEXT NOT NULL DEFAULT (datetime('now')),
+  updatedAt TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE UNIQUE INDEX idx_scripts_name_scope ON scripts(name, scope, COALESCE(scopeId, ''));
+CREATE INDEX idx_scripts_scope ON scripts(scope, scopeId);
+CREATE INDEX idx_scripts_scratch ON scripts(isScratch, createdAt);
+
+CREATE TABLE script_versions (
+  id TEXT PRIMARY KEY,
+  scriptId TEXT NOT NULL REFERENCES scripts(id) ON DELETE CASCADE,
+  version INTEGER NOT NULL,
+  source TEXT NOT NULL,
+  description TEXT NOT NULL,
+  intent TEXT NOT NULL,
+  signatureJson TEXT NOT NULL,
+  contentHash TEXT NOT NULL,
+  changedByAgentId TEXT,
+  changedAt TEXT NOT NULL DEFAULT (datetime('now')),
+  changeReason TEXT,
+  UNIQUE(scriptId, version)
+);
+
+CREATE INDEX idx_script_versions_hash ON script_versions(contentHash);

package/src/be/migrations/065_script_embeddings.sql

@@ -0,0 +1,7 @@
+CREATE TABLE script_embeddings (
+  scriptId TEXT PRIMARY KEY REFERENCES scripts(id) ON DELETE CASCADE,
+  embedding BLOB NOT NULL,
+  embeddingModel TEXT NOT NULL,
+  embeddedText TEXT NOT NULL,
+  embeddedAt TEXT NOT NULL DEFAULT (datetime('now'))
+);

package/src/be/pricing-normalize.ts

@@ -0,0 +1,81 @@
+/**
+ * Phase 2 fix — normalize provider model ids before pricing-table lookup.
+ *
+ * Different harnesses report the same underlying model under different keys:
+ *
+ *   - claude-adapter      → `claude-opus-4-7`               (bare)
+ *   - codex-adapter       → `gpt-5.4`                       (bare, dotted)
+ *   - opencode-adapter    → `openrouter/anthropic/claude-sonnet-4.5`
+ *   - pi-mono-adapter     → `github-copilot/gpt-5.4` or
+ *                            `openrouter/anthropic/claude-sonnet-4.5`
+ *
+ * The pricing seed in `src/be/seed-pricing.ts` keys by what models.dev calls
+ * the model (e.g. `anthropic/claude-sonnet-4.5` for openrouter rows,
+ * `gpt-5.4` for openai rows). That means harness-emitted ids with extra
+ * routing prefixes (`openrouter/`, `github-copilot/`, …) fall through to
+ * `costSource='unpriced'` even when we have a perfectly good rate row.
+ *
+ * Rather than rewriting the adapter outputs (which are the harness's source
+ * of truth and useful for debugging), we normalize at the *lookup boundary*:
+ * strip noisy routing prefixes so the seeded canonical key resolves.
+ *
+ * Apply this helper symmetrically: once when seeding rows (so seed keys are
+ * canonical) and once when querying (so adapter-emitted keys collapse onto
+ * the same canonical form).
+ */
+
+import type { PricingProvider } from "../types";
+
+/**
+ * Routing prefixes that a harness may prepend to the underlying model id but
+ * that have no pricing semantics. Stripping these collapses
+ * `openrouter/anthropic/claude-sonnet-4.5` → `anthropic/claude-sonnet-4.5`
+ * which is the key models.dev/openrouter uses.
+ *
+ * Order matters: we only ever strip the *first* matching prefix so we don't
+ * accidentally chew through a model id like `openai/openai-test-model`.
+ */
+const ROUTING_PREFIXES_BY_PROVIDER: Record<PricingProvider, readonly string[]> = {
+  // opencode routes via opencode-server which proxies to openrouter, anthropic,
+  // openai, … — strip whichever proxy prefix the user picked.
+  opencode: ["openrouter/", "github-copilot/"],
+  // pi-mono can hit openrouter mirrors, the github-copilot proxy, or native
+  // anthropic/openai/google providers.
+  pi: ["openrouter/", "github-copilot/"],
+  // codex normally reports a bare id, but a user may set MODEL_OVERRIDE to a
+  // prefixed form. Be forgiving on the lookup side.
+  codex: ["openai/", "github-copilot/"],
+  // claude / claude-managed / devin / gemini emit bare ids today. The empty
+  // list keeps the helper a no-op for them but the entry-per-provider shape
+  // means a future provider can opt in without changing call-sites.
+  claude: [],
+  "claude-managed": [],
+  devin: [],
+  gemini: [],
+};
+
+/**
+ * Canonical model key for a `(provider, model)` pair. Idempotent — calling
+ * this on an already-normalized value is a no-op.
+ *
+ * Rules:
+ *  1. Lowercase the input. Adapters sometimes pass mixed case (codex calls
+ *     `.toLowerCase()` itself; opencode/pi don't always).
+ *  2. Strip the first matching routing prefix for this provider, if any.
+ *
+ * We deliberately do NOT touch dotted-vs-dashed minor versions
+ * (`gpt-5.4` vs `gpt-5-4`) — both harness output and models.dev use dotted
+ * for openai and dashed for anthropic, so there's no real drift there.
+ */
+export function normalizeModelKey(provider: PricingProvider, model: string): string {
+  if (!model) return model;
+  let key = model.toLowerCase();
+  const prefixes = ROUTING_PREFIXES_BY_PROVIDER[provider] ?? [];
+  for (const prefix of prefixes) {
+    if (key.startsWith(prefix)) {
+      key = key.slice(prefix.length);
+      break;
+    }
+  }
+  return key;
+}