stagent 0.1.10 → 0.1.11

package/src/lib/agents/learned-context.ts

@@ -2,7 +2,7 @@ import { db } from "@/lib/db";
 import { learnedContext, notifications } from "@/lib/db/schema";
 import { and, desc, eq } from "drizzle-orm";
 import type { LearnedContextRow } from "@/lib/db/schema";
-import Anthropic from "@anthropic-ai/sdk";
+import { runMetaCompletion } from "./runtime/claude";
 
 const CONTEXT_CHAR_LIMIT = 8_000;
 const SUMMARIZATION_THRESHOLD = 6_000;
@@ -243,14 +243,8 @@ export async function summarizeContext(profileId: string): Promise<void> {
   const content = getActiveLearnedContext(profileId);
   if (!content || content.length <= SUMMARIZATION_THRESHOLD) return;
 
-  const client = new Anthropic();
-  const response = await client.messages.create({
-    model: "claude-sonnet-4-20250514",
-    max_tokens: 2048,
-    messages: [
-      {
-        role: "user",
-        content: `You are condensing learned context for an AI agent profile "${profileId}".
+  const { text } = await runMetaCompletion({
+    prompt: `You are condensing learned context for an AI agent profile "${profileId}".
 The current context has grown to ${content.length} characters and needs to be summarized to under ${SUMMARIZATION_THRESHOLD} characters while preserving all key patterns, best practices, and important insights.
 
 Current learned context:
@@ -266,12 +260,10 @@ Produce a condensed version that:
 5. Stays under ${SUMMARIZATION_THRESHOLD} characters
 
 Output ONLY the condensed context, no preamble.`,
-      },
-    ],
+    activityType: "context_summarization",
   });
 
-  const summarized =
-    response.content[0].type === "text" ? response.content[0].text : "";
+  const summarized = text.trim();
 
   if (!summarized || summarized.length >= content.length) return;
 

package/src/lib/agents/pattern-extractor.ts

@@ -1,4 +1,3 @@
-import Anthropic from "@anthropic-ai/sdk";
 import { db } from "@/lib/db";
 import { tasks, agentLogs } from "@/lib/db/schema";
 import { eq, desc } from "drizzle-orm";
@@ -6,6 +5,7 @@ import {
   getActiveLearnedContext,
   proposeContextAddition,
 } from "./learned-context";
+import { runMetaCompletion } from "./runtime/claude";
 
 export interface PatternEntry {
   title: string;
@@ -17,50 +17,9 @@ export interface PatternProposal {
   patterns: PatternEntry[];
 }
 
-const PATTERN_TOOL: Anthropic.Messages.Tool = {
-  name: "propose_learned_patterns",
-  description:
-    "Propose patterns learned from this task execution that should be remembered for future tasks with this profile.",
-  input_schema: {
-    type: "object" as const,
-    properties: {
-      patterns: {
-        type: "array",
-        items: {
-          type: "object",
-          properties: {
-            title: {
-              type: "string",
-              description: "Short pattern name (2-6 words)",
-            },
-            description: {
-              type: "string",
-              description:
-                "Concise description of the pattern or lesson (1-2 sentences)",
-            },
-            category: {
-              type: "string",
-              enum: [
-                "error_resolution",
-                "best_practice",
-                "shortcut",
-                "preference",
-              ],
-            },
-          },
-          required: ["title", "description", "category"],
-        },
-        description:
-          "Patterns worth remembering. Return empty array if nothing notable.",
-      },
-    },
-    required: ["patterns"],
-  },
-};
-
 /**
  * Analyze a completed task for patterns worth learning.
- * Makes a focused Claude API call, then proposes additions if patterns found.
+ * Routes through the Claude Agent SDK runtime (no direct Anthropic SDK usage).
  * Returns the notification ID if a proposal was created, null otherwise.
  */
 export async function analyzeForLearnedPatterns(
@@ -99,16 +58,13 @@ export async function analyzeForLearnedPatterns(
     })
     .join("\n");
 
-  const client = new Anthropic();
-  const response = await client.messages.create({
-    model: "claude-sonnet-4-20250514",
-    max_tokens: 1024,
-    tools: [PATTERN_TOOL],
-    tool_choice: { type: "tool", name: "propose_learned_patterns" },
-    messages: [
-      {
-        role: "user",
-        content: `Analyze this completed task for patterns worth learning for the "${profileId}" agent profile.
+  const { text } = await runMetaCompletion({
+    prompt: `Analyze this completed task for patterns worth learning for the "${profileId}" agent profile.
+
+Return ONLY a JSON array (no markdown, no code fences):
+[{"title": "...", "description": "...", "category": "error_resolution|best_practice|shortcut|preference"}]
+
+Return an empty array [] if no noteworthy patterns.
 
 ## Task
 Title: ${task.title}
@@ -124,22 +80,17 @@ ${logSummary.slice(0, 2000)}
 ${currentContext ?? "(none yet)"}
 
 Extract ONLY genuinely useful patterns — things that would help this profile avoid mistakes or work more efficiently on similar future tasks. If this task was routine with nothing notable, return an empty patterns array. Do NOT repeat patterns already in the learned context.`,
-      },
-    ],
+    activityType: "pattern_extraction",
   });
 
-  // Extract the tool use result
-  const toolBlock = response.content.find(
-    (block) => block.type === "tool_use" && block.name === "propose_learned_patterns"
-  );
-
-  if (!toolBlock || toolBlock.type !== "tool_use") return null;
+  // Parse JSON array from response text
+  const jsonMatch = text.match(/\[[\s\S]*\]/);
+  const patterns: PatternEntry[] = jsonMatch ? JSON.parse(jsonMatch[0]) : [];
 
-  const proposal = toolBlock.input as PatternProposal;
-  if (!proposal.patterns || proposal.patterns.length === 0) return null;
+  if (patterns.length === 0) return null;
 
   // Format patterns as text for the proposal
-  const formattedAdditions = proposal.patterns
+  const formattedAdditions = patterns
     .map(
       (p) =>
         `### ${p.title} [${p.category}]\n${p.description}`

package/src/lib/agents/profiles/builtins/code-reviewer/profile.yaml

@@ -15,7 +15,6 @@ canUseToolPolicy:
   autoApprove: [Read, Grep, Glob]
   autoDeny: []
 
-temperature: 0.3
 maxTurns: 20
 outputFormat: structured-findings
 

package/src/lib/agents/profiles/builtins/data-analyst/profile.yaml

@@ -15,7 +15,6 @@ canUseToolPolicy:
   autoApprove: [Read, Grep, Glob]
   autoDeny: []
 
-temperature: 0.3
 maxTurns: 30
 
 author: stagent

package/src/lib/agents/profiles/builtins/devops-engineer/profile.yaml

@@ -15,7 +15,6 @@ canUseToolPolicy:
   autoApprove: [Read, Grep, Glob]
   autoDeny: []
 
-temperature: 0.3
 maxTurns: 30
 
 author: stagent

package/src/lib/agents/profiles/builtins/document-writer/profile.yaml

@@ -14,7 +14,6 @@ canUseToolPolicy:
   autoApprove: [Read]
   autoDeny: []
 
-temperature: 0.5
 maxTurns: 20
 outputFormat: markdown-document
 

package/src/lib/agents/profiles/builtins/general/profile.yaml

@@ -12,7 +12,6 @@ runtimeOverrides:
       Stay pragmatic, execute the requested work directly, and prefer concise operational updates.
       Keep outputs grounded in the current workspace and call out blocked actions explicitly.
 
-temperature: 0.5
 maxTurns: 30
 
 author: stagent

package/src/lib/agents/profiles/builtins/health-fitness-coach/profile.yaml

@@ -14,7 +14,6 @@ canUseToolPolicy:
   autoApprove: [WebSearch, WebFetch, Read]
   autoDeny: [Bash, Write, Edit]
 
-temperature: 0.6
 maxTurns: 20
 
 author: stagent

package/src/lib/agents/profiles/builtins/learning-coach/profile.yaml

@@ -14,7 +14,6 @@ canUseToolPolicy:
   autoApprove: [WebSearch, WebFetch, Read]
   autoDeny: [Bash, Write, Edit]
 
-temperature: 0.5
 maxTurns: 25
 
 author: stagent

package/src/lib/agents/profiles/builtins/project-manager/profile.yaml

@@ -14,7 +14,6 @@ canUseToolPolicy:
   autoApprove: [Read, Grep, Glob]
   autoDeny: []
 
-temperature: 0.4
 maxTurns: 25
 
 author: stagent

package/src/lib/agents/profiles/builtins/researcher/profile.yaml

@@ -15,7 +15,6 @@ canUseToolPolicy:
   autoApprove: [WebSearch, WebFetch, Read]
   autoDeny: []
 
-temperature: 0.4
 maxTurns: 25
 
 author: stagent

package/src/lib/agents/profiles/builtins/shopping-assistant/profile.yaml

@@ -14,7 +14,6 @@ canUseToolPolicy:
   autoApprove: [WebSearch, WebFetch, Read]
   autoDeny: [Bash, Write, Edit]
 
-temperature: 0.5
 maxTurns: 20
 
 author: stagent

package/src/lib/agents/profiles/builtins/sweep/profile.yaml

@@ -5,7 +5,6 @@ domain: work
 tags: [sweep, audit, improvement, maintenance]
 supportedRuntimes: [claude-code]
 
-temperature: 0.3
 maxTurns: 50
 outputFormat: json
 

package/src/lib/agents/profiles/builtins/technical-writer/profile.yaml

@@ -16,7 +16,6 @@ canUseToolPolicy:
   autoApprove: [Read, Grep, Glob]
   autoDeny: []
 
-temperature: 0.4
 maxTurns: 20
 outputFormat: markdown
 

package/src/lib/agents/profiles/builtins/travel-planner/profile.yaml

@@ -14,7 +14,6 @@ canUseToolPolicy:
   autoApprove: [WebSearch, WebFetch, Read]
   autoDeny: [Bash, Write, Edit]
 
-temperature: 0.6
 maxTurns: 25
 
 author: stagent

package/src/lib/agents/profiles/builtins/wealth-manager/profile.yaml

@@ -14,7 +14,6 @@ canUseToolPolicy:
   autoApprove: [Read]
   autoDeny: [Bash, Write, Edit]
 
-temperature: 0.3
 maxTurns: 20
 
 author: stagent

package/src/lib/agents/profiles/registry.ts

@@ -174,7 +174,6 @@ function scanProfiles(): Map<string, AgentProfile> {
         allowedTools: config.allowedTools,
         mcpServers: config.mcpServers as Record<string, unknown>,
         canUseToolPolicy: config.canUseToolPolicy,
-        temperature: config.temperature,
         maxTurns: config.maxTurns,
         outputFormat: config.outputFormat,
         version: config.version,

package/src/lib/agents/profiles/types.ts

@@ -31,7 +31,6 @@ export interface AgentProfile {
   allowedTools?: string[];
   mcpServers?: Record<string, unknown>;
   canUseToolPolicy?: CanUseToolPolicy;
-  temperature?: number;
   maxTurns?: number;
   outputFormat?: string;
   version?: string;

package/src/lib/agents/runtime/catalog.ts

@@ -50,7 +50,7 @@ const RUNTIME_CATALOG: Record<AgentRuntimeId, RuntimeCatalogEntry> = {
       resume: true,
       cancel: true,
       approvals: true,
-      mcpServers: true,
+      mcpServers: false, // Not yet wired — configs not passed to codex subprocess
       profileTests: false,
       taskAssist: true,
       authHealthCheck: true,

package/src/lib/agents/runtime/claude.ts

@@ -252,6 +252,71 @@ async function runClaudeProfileTests(profileId: string): Promise<ProfileTestRepo
   };
 }
 
+// ---------------------------------------------------------------------------
+// Lightweight meta-completion (pattern extraction, context summarization, etc.)
+// ---------------------------------------------------------------------------
+
+export async function runMetaCompletion(input: {
+  prompt: string;
+  activityType: string;
+}): Promise<{ text: string; usage: UsageSnapshot }> {
+  const authEnv = await getAuthEnv();
+  const startedAt = new Date();
+  let usage: UsageSnapshot = {};
+  const abortController = new AbortController();
+  const timeout = setTimeout(() => abortController.abort(), 60_000);
+
+  try {
+    const response = query({
+      prompt: input.prompt,
+      options: {
+        abortController,
+        includePartialMessages: true,
+        cwd: process.cwd(),
+        env: buildClaudeSdkEnv(authEnv),
+        allowedTools: [],
+        maxTurns: 1,
+      },
+    });
+
+    const collected = await collectResultText(
+      response as AsyncIterable<Record<string, unknown>>
+    );
+    usage = collected.usage;
+
+    await recordUsageLedgerEntry({
+      activityType: input.activityType as import("@/lib/usage/ledger").UsageActivityType,
+      runtimeId: "claude-code",
+      providerId: "anthropic",
+      modelId: usage.modelId ?? null,
+      inputTokens: usage.inputTokens ?? null,
+      outputTokens: usage.outputTokens ?? null,
+      totalTokens: usage.totalTokens ?? null,
+      status: "completed",
+      startedAt,
+      finishedAt: new Date(),
+    });
+
+    return { text: collected.resultText, usage };
+  } catch (error) {
+    await recordUsageLedgerEntry({
+      activityType: input.activityType as import("@/lib/usage/ledger").UsageActivityType,
+      runtimeId: "claude-code",
+      providerId: "anthropic",
+      modelId: usage.modelId ?? null,
+      inputTokens: usage.inputTokens ?? null,
+      outputTokens: usage.outputTokens ?? null,
+      totalTokens: usage.totalTokens ?? null,
+      status: "failed",
+      startedAt,
+      finishedAt: new Date(),
+    });
+    throw error;
+  } finally {
+    clearTimeout(timeout);
+  }
+}
+
 async function runClaudeTaskAssist(
   input: TaskAssistInput
 ): Promise<TaskAssistResponse> {
@@ -345,6 +410,7 @@ async function testClaudeConnection(): Promise<RuntimeConnectionResult> {
       options: {
         abortController,
         maxTurns: 1,
+        includePartialMessages: false,
         cwd: process.cwd(),
         env: buildClaudeSdkEnv(authEnv),
       },

package/src/lib/constants/task-status.ts

@@ -47,3 +47,9 @@ export function isValidDragTransition(from: TaskStatus, to: TaskStatus): boolean
 
 /** Maximum number of times a task can be resumed before requiring a fresh start */
 export const MAX_RESUME_COUNT = 3;
+
+/** Default max turns for agent task execution (safety net) */
+export const DEFAULT_MAX_TURNS = 50;
+
+/** Default per-execution budget cap in USD */
+export const DEFAULT_MAX_BUDGET_USD = 2.0;

package/src/lib/data/seed-data/profiles.ts

@@ -34,7 +34,6 @@ export function getSampleProfiles(): SampleProfileSeed[] {
         canUseToolPolicy: {
           autoApprove: ["Read", "Grep"],
         },
-        temperature: 0.3,
         maxTurns: 18,
         outputFormat: "Weekly operating note with metrics, risks, and next actions.",
         author: SAMPLE_PROFILE_AUTHOR,
@@ -73,7 +72,6 @@ You review pipeline movement, funnel risk, and rep follow-ups with a bias toward
         canUseToolPolicy: {
           autoApprove: ["Read"],
         },
-        temperature: 0.6,
         maxTurns: 16,
         outputFormat: "Experiment summary with winning message angles and next tests.",
         author: SAMPLE_PROFILE_AUTHOR,
@@ -109,7 +107,6 @@ You turn campaign performance and research inputs into sharper launch messaging.
         domain: "personal",
         tags: ["investing", "portfolio", "risk", "habits"],
         allowedTools: ["Read", "Write"],
-        temperature: 0.25,
         maxTurns: 14,
         outputFormat: "Short investor brief with posture, risk notes, and watchlist changes.",
         author: SAMPLE_PROFILE_AUTHOR,

package/src/lib/usage/__tests__/ledger.test.ts

@@ -33,9 +33,10 @@ function formatLocalDay(date: Date) {
 }
 
 describe("usage ledger", () => {
-  it("records normalized ledger rows with derived and unknown pricing states", async () => {
+  it("records normalized ledger rows with derived, fallback, and unknown pricing states", async () => {
     const { db, usageLedger, recordUsageLedgerEntry } = await loadUsageModules();
 
+    // Known model — gets specific pricing rule
     await recordUsageLedgerEntry({
       activityType: "task_assist",
       runtimeId: "claude-code",
@@ -49,6 +50,7 @@ describe("usage ledger", () => {
       finishedAt: new Date("2026-03-10T08:01:00.000Z"),
     });
 
+    // Unknown model — hits catch-all fallback pricing (conservative estimate)
     await recordUsageLedgerEntry({
       activityType: "task_assist",
       runtimeId: "openai-codex-app-server",
@@ -62,15 +64,37 @@ describe("usage ledger", () => {
       finishedAt: new Date("2026-03-10T09:01:00.000Z"),
     });
 
+    // Null modelId — gets unknown_pricing (no model to match)
+    await recordUsageLedgerEntry({
+      activityType: "task_run",
+      runtimeId: "claude-code",
+      providerId: "anthropic",
+      modelId: null,
+      inputTokens: 100,
+      outputTokens: 50,
+      totalTokens: 150,
+      status: "completed",
+      startedAt: new Date("2026-03-10T10:00:00.000Z"),
+      finishedAt: new Date("2026-03-10T10:01:00.000Z"),
+    });
+
     const rows = await db.select().from(usageLedger);
-    expect(rows).toHaveLength(2);
+    expect(rows).toHaveLength(3);
 
-    const priced = rows.find((row) => row.providerId === "anthropic");
+    // Known: specific pricing
+    const priced = rows.find((row) => row.modelId === "claude-sonnet-4-20250514");
     expect(priced?.costMicros).toBe(10_500);
     expect(priced?.status).toBe("completed");
-    expect(priced?.pricingVersion).toBe("registry-2026-03-12");
+    expect(priced?.pricingVersion).toBe("registry-2026-03-15");
+
+    // Unknown model: fallback pricing (conservative Opus-tier for OpenAI: $10/$30)
+    const fallback = rows.find((row) => row.modelId === "codex-unknown");
+    expect(fallback?.costMicros).toBeGreaterThan(0);
+    expect(fallback?.status).toBe("completed");
+    expect(fallback?.pricingVersion).toBe("registry-2026-03-15-fallback");
 
-    const unknown = rows.find((row) => row.providerId === "openai");
+    // Null modelId: truly unknown
+    const unknown = rows.find((row) => row.modelId === null);
     expect(unknown?.costMicros).toBeNull();
     expect(unknown?.status).toBe("unknown_pricing");
     expect(unknown?.pricingVersion).toBeNull();

package/src/lib/usage/ledger.ts

@@ -15,7 +15,9 @@ export type UsageActivityType =
   | "workflow_step"
   | "scheduled_firing"
   | "task_assist"
-  | "profile_test";
+  | "profile_test"
+  | "pattern_extraction"
+  | "context_summarization";
 
 export type UsageLedgerStatus =
   | "completed"

package/src/lib/usage/pricing.ts

@@ -7,25 +7,79 @@ export interface PricingRule {
 }
 
 const PRICING_RULES: PricingRule[] = [
+  // ── Anthropic ──────────────────────────────────────────────────────
   {
     providerId: "anthropic",
-    pricingVersion: "registry-2026-03-12",
+    pricingVersion: "registry-2026-03-15",
+    inputCostPerMillionMicros: 15_000_000,
+    outputCostPerMillionMicros: 75_000_000,
+    matchesModel(modelId) {
+      return modelId.startsWith("claude-opus");
+    },
+  },
+  {
+    providerId: "anthropic",
+    pricingVersion: "registry-2026-03-15",
     inputCostPerMillionMicros: 3_000_000,
     outputCostPerMillionMicros: 15_000_000,
     matchesModel(modelId) {
-      return (
-        modelId === "claude-sonnet-4-20250514" ||
-        modelId.startsWith("claude-sonnet-4")
-      );
+      return modelId.startsWith("claude-sonnet");
+    },
+  },
+  {
+    providerId: "anthropic",
+    pricingVersion: "registry-2026-03-15",
+    inputCostPerMillionMicros: 800_000,
+    outputCostPerMillionMicros: 4_000_000,
+    matchesModel(modelId) {
+      return modelId.startsWith("claude-haiku");
     },
   },
+  // ── OpenAI ─────────────────────────────────────────────────────────
   {
     providerId: "openai",
-    pricingVersion: "registry-2026-03-12",
+    pricingVersion: "registry-2026-03-15",
     inputCostPerMillionMicros: 1_500_000,
     outputCostPerMillionMicros: 6_000_000,
     matchesModel(modelId) {
-      return modelId === "codex-mini-latest" || modelId.startsWith("codex-mini");
+      return modelId.startsWith("codex-mini") || modelId === "codex-mini-latest";
+    },
+  },
+  {
+    providerId: "openai",
+    pricingVersion: "registry-2026-03-15",
+    inputCostPerMillionMicros: 2_500_000,
+    outputCostPerMillionMicros: 10_000_000,
+    matchesModel(modelId) {
+      return modelId.startsWith("gpt-4o");
+    },
+  },
+  {
+    providerId: "openai",
+    pricingVersion: "registry-2026-03-15",
+    inputCostPerMillionMicros: 10_000_000,
+    outputCostPerMillionMicros: 30_000_000,
+    matchesModel(modelId) {
+      return modelId.startsWith("gpt-5") || modelId.startsWith("o3") || modelId.startsWith("o4");
+    },
+  },
+  // ── Catch-all (conservative estimate to prevent null costs) ────────
+  {
+    providerId: "anthropic",
+    pricingVersion: "registry-2026-03-15-fallback",
+    inputCostPerMillionMicros: 15_000_000,
+    outputCostPerMillionMicros: 75_000_000,
+    matchesModel() {
+      return true;
+    },
+  },
+  {
+    providerId: "openai",
+    pricingVersion: "registry-2026-03-15-fallback",
+    inputCostPerMillionMicros: 10_000_000,
+    outputCostPerMillionMicros: 30_000_000,
+    matchesModel() {
+      return true;
     },
   },
 ];

package/src/lib/validators/__tests__/profile.test.ts

@@ -28,7 +28,6 @@ describe("ProfileConfigSchema", () => {
         preToolCall: ["echo pre"],
         postToolCall: ["echo post"],
       },
-      temperature: 0.5,
       maxTurns: 20,
       outputFormat: "markdown",
       author: "stagent",
@@ -77,20 +76,6 @@ describe("ProfileConfigSchema", () => {
     expect(result.success).toBe(false);
   });
 
-  it("rejects temperature out of range", () => {
-    const tooHigh = ProfileConfigSchema.safeParse({
-      ...validProfile,
-      temperature: 1.5,
-    });
-    expect(tooHigh.success).toBe(false);
-
-    const tooLow = ProfileConfigSchema.safeParse({
-      ...validProfile,
-      temperature: -0.1,
-    });
-    expect(tooLow.success).toBe(false);
-  });
-
   it("rejects invalid source URL", () => {
     const result = ProfileConfigSchema.safeParse({
       ...validProfile,

package/src/lib/validators/profile.ts

@@ -38,7 +38,6 @@ export const ProfileConfigSchema = z.object({
       postToolCall: z.array(z.string()).optional(),
     })
     .optional(),
-  temperature: z.number().min(0).max(1).optional(),
   maxTurns: z.number().positive().optional(),
   outputFormat: z.string().optional(),
   author: z.string().optional(),

package/src/lib/workflows/__tests__/engine.test.ts

@@ -103,6 +103,8 @@ describe("executeWorkflow", () => {
       .mockResolvedValueOnce([workflow])
       .mockResolvedValueOnce([failedTask])
       .mockResolvedValueOnce([workflow])
+      .mockResolvedValueOnce([workflow])
+      // syncSourceTaskStatus reads the workflow to find sourceTaskId
       .mockResolvedValueOnce([workflow]);
 
     const { executeWorkflow } = await import("../engine");

package/src/lib/workflows/engine.ts

@@ -837,11 +837,12 @@ async function syncSourceTaskStatus(
   status: "completed" | "failed"
 ): Promise<void> {
   try {
-    const [workflow] = await db
+    const result = await db
       .select()
       .from(workflows)
       .where(eq(workflows.id, workflowId));
 
+    const workflow = Array.isArray(result) ? result[0] : undefined;
     if (!workflow) return;
 
     const def = JSON.parse(workflow.definition);