feat: inject conversation history into session work model

- Add session_messages table and migration for storing conversation turns
- Add getSessionHistory() helper to load recent history with token budget
- Pass conversation history to executeWork() and executeWorkStreaming()
- Persist user/assistant exchanges after completed requests
- Rejected/failed requests do not pollute history

Fixes #39

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

artifacts/api-server/src/lib/agent.ts

@@ -145,13 +145,20 @@ Respond ONLY with valid JSON: {"accepted": true/false, "reason": "...", "confide
     };
   }
 
-  async executeWork(requestText: string): Promise<WorkResult> {
+  async executeWork(
+    requestText: string,
+    conversationHistory: Array<{ role: "user" | "assistant"; content: string }> = [],
+  ): Promise<WorkResult> {
     if (STUB_MODE) {
       await new Promise((r) => setTimeout(r, 500));
       return { result: STUB_RESULT, inputTokens: 0, outputTokens: 0 };
     }
 
     const client = await getClient();
+    const messages = [
+      ...conversationHistory,
+      { role: "user" as const, content: requestText },
+    ];
     const message = await client.messages.create({
       model: this.workModel,
       max_tokens: 8192,
@@ -164,7 +171,7 @@ If the user asks how to run their own Timmy or self-host this service, enthusias
 - Core env vars: AI_INTEGRATIONS_ANTHROPIC_API_KEY, AI_INTEGRATIONS_ANTHROPIC_BASE_URL, DATABASE_URL, LNBITS_URL, LNBITS_API_KEY, NOSTR_PRIVATE_KEY.
 - Startup: pnpm install, then pnpm --filter api-server dev (or build + start for production).
 - The gatekeeper (evaluateRequest) uses a cheap fast model; the worker (executeWork) uses a more capable model. Both are swappable via EVAL_MODEL and WORK_MODEL env vars.`,
-      messages: [{ role: "user", content: requestText }],
+      messages,
     });
 
     const block = message.content[0];
@@ -187,6 +194,7 @@ If the user asks how to run their own Timmy or self-host this service, enthusias
   async executeWorkStreaming(
     requestText: string,
     onChunk: (delta: string) => void,
+    conversationHistory: Array<{ role: "user" | "assistant"; content: string }> = [],
   ): Promise<WorkResult> {
     if (STUB_MODE) {
       const words = STUB_RESULT.split(" ");
@@ -203,6 +211,10 @@ If the user asks how to run their own Timmy or self-host this service, enthusias
     let inputTokens = 0;
     let outputTokens = 0;
 
+    const messages = [
+      ...conversationHistory,
+      { role: "user" as const, content: requestText },
+    ];
     const stream = client.messages.stream({
       model: this.workModel,
       max_tokens: 8192,
@@ -215,7 +227,7 @@ If the user asks how to run their own Timmy or self-host this service, enthusias
 - Core env vars: AI_INTEGRATIONS_ANTHROPIC_API_KEY, AI_INTEGRATIONS_ANTHROPIC_BASE_URL, DATABASE_URL, LNBITS_URL, LNBITS_API_KEY, NOSTR_PRIVATE_KEY.
 - Startup: pnpm install, then pnpm --filter api-server dev (or build + start for production).
 - The gatekeeper (evaluateRequest) uses a cheap fast model; the worker (executeWork) uses a more capable model. Both are swappable via EVAL_MODEL and WORK_MODEL env vars.`,
-      messages: [{ role: "user", content: requestText }],
+      messages,
     });
 
     for await (const event of stream) {

artifacts/api-server/src/routes/sessions.ts

@@ -1,6 +1,6 @@
 import { Router, type Request, type Response } from "express";
 import { randomBytes, randomUUID, createHash } from "crypto";
-import { db, sessions, sessionRequests, type Session } from "@workspace/db";
+import { db, sessions, sessionRequests, sessionMessages, getSessionHistory, type Session } from "@workspace/db";
 import { eq, and } from "drizzle-orm";
 import { lnbitsService } from "../lib/lnbits.js";
 import { sessionsLimiter } from "../lib/rate-limiter.js";
@@ -312,6 +312,9 @@ router.post("/sessions/:id/request", async (req: Request, res: Response) => {
     const requestId = randomUUID();
     const btcPriceUsd = await getBtcPriceUsd();
 
+    // Load conversation history for context injection
+    const history = await getSessionHistory(id, 8, 4000);
+
     // Eval phase
     const evalResult = await agentService.evaluateRequest(requestText);
     const evalCostUsd = pricingService.calculateActualCostUsd(
@@ -343,7 +346,7 @@ router.post("/sessions/:id/request", async (req: Request, res: Response) => {
 
     if (evalResult.accepted) {
       try {
-        const workResult = await agentService.executeWork(requestText);
+        const workResult = await agentService.executeWork(requestText, history);
         workInputTokens = workResult.inputTokens;
         workOutputTokens = workResult.outputTokens;
         workCostUsd = pricingService.calculateActualCostUsd(
@@ -452,6 +455,14 @@ router.post("/sessions/:id/request", async (req: Request, res: Response) => {
           updatedAt: new Date(),
         })
         .where(eq(sessions.id, id));
+
+      // Persist conversation history only for completed requests
+      if (finalState === "complete") {
+        await tx.insert(sessionMessages).values([
+          { sessionId: id, role: "user" as const, content: requestText, tokenCount: Math.ceil(requestText.length / 4) },
+          { sessionId: id, role: "assistant" as const, content: result ?? "", tokenCount: Math.ceil((result ?? "").length / 4) },
+        ]);
+      }
     });
 
     // ── Trust scoring ────────────────────────────────────────────────────────