daemora 1.0.5 → 1.0.7

package/src/{systemPrompt.js → agents/systemPrompt.js}

@@ -1,12 +1,40 @@
 import { readFileSync, existsSync } from "fs";
 import { join } from "path";
-import { config } from "./config/default.js";
-import skillLoader from "./skills/SkillLoader.js";
-import mcpManager from "./mcp/MCPManager.js";
-import tenantContext from "./tenants/TenantContext.js";
+import { config } from "../config/default.js";
+import skillLoader from "../skills/SkillLoader.js";
+import mcpManager from "../mcp/MCPManager.js";
+import tenantContext from "../tenants/TenantContext.js";
+
+// ── Tool → required env keys mapping ──────────────────────────────────────────
+// Tools listed here need at least ONE of their required keys set.
+// Unconfigured tools are excluded from full docs and listed as [NO AUTH].
+const TOOL_REQUIRED_KEYS = {
+  sendEmail:       ["RESEND_API_KEY", "EMAIL_USER"],
+  makeVoiceCall:   ["TWILIO_ACCOUNT_SID"],
+  transcribeAudio: ["OPENAI_API_KEY"],
+  textToSpeech:    ["OPENAI_API_KEY", "ELEVENLABS_API_KEY"],
+  generateImage:   ["OPENAI_API_KEY"],
+  googlePlaces:    ["GOOGLE_PLACES_API_KEY"],
+  calendar:        ["GOOGLE_CALENDAR_API_KEY"],
+  contacts:        ["GOOGLE_CONTACTS_ACCESS_TOKEN"],
+  philipsHue:      ["HUE_BRIDGE_IP"],
+  sonos:           ["SONOS_HOST"],
+  database:        ["DATABASE_URL", "MYSQL_URL"],
+  sshTool:         ["SSH_DEFAULT_HOST"],
+};
+
+function _getConfiguredKeys() {
+  const store = tenantContext.getStore();
+  const tenantKeys = store?.apiKeys || {};
+  return { ...process.env, ...tenantKeys };
+}
 
-skillLoader.load();
-skillLoader.embedSkills().catch(() => {});  // Pre-compute skill embeddings at startup (fire-and-forget)
+function _isToolConfigured(toolName) {
+  const requiredKeys = TOOL_REQUIRED_KEYS[toolName];
+  if (!requiredKeys) return true;
+  const env = _getConfiguredKeys();
+  return requiredKeys.some(key => !!env[key]);
+}
 
 /**
  * Build the system prompt dynamically by composing modular sections.
@@ -15,19 +43,20 @@ skillLoader.embedSkills().catch(() => {});  // Pre-compute skill embeddings at s
  * @param {object} [runtimeMeta] - Optional metadata for runtime line { model, agentId, thinkingLevel }
  */
 export async function buildSystemPrompt(taskInput, promptMode = "full", runtimeMeta = {}) {
-  // Minimal mode: Soul + ResponseFormat + ToolDocs + MCP + SubagentContext only
-  // Skips: Memory, DailyLog, SemanticRecall, Skills, OperationalGuidelines
   const sections = promptMode === "minimal"
     ? await Promise.all([
         renderSoul(),
+        renderUserProfile(),
         renderResponseFormat(),
         renderToolDocs(),
         renderMCPTools(),
         renderToolUsageRules(),
+        renderSkills(taskInput, 10),
         renderSubagentContext(runtimeMeta.taskDescription || taskInput),
       ])
     : await Promise.all([
         renderSoul(),
+        renderUserProfile(),
         renderResponseFormat(),
         renderToolDocs(),
         renderMCPTools(),
@@ -39,7 +68,6 @@ export async function buildSystemPrompt(taskInput, promptMode = "full", runtimeM
         renderOperationalGuidelines(),
       ]);
 
-  // Always append runtime line
   const runtime = renderRuntime(runtimeMeta);
   if (runtime) sections.push(runtime);
 
@@ -49,8 +77,7 @@ export async function buildSystemPrompt(taskInput, promptMode = "full", runtimeM
   };
 }
 
-// ── Tenant-aware memory path resolution ───────────────────────────────────────
-// Called at render time so TenantContext is active (we're inside tenantContext.run(...)).
+// ── Tenant-aware path resolution ─────────────────────────────────────────────
 
 function _getContextMemoryPaths() {
   const store = tenantContext.getStore();
@@ -63,15 +90,10 @@ function _getContextMemoryPaths() {
   return { memoryPath: config.memoryPath, memoryDir: config.memoryDir, tenantId: null };
 }
 
-/**
- * Inject the top-k most semantically relevant memories for this specific task.
- * Only runs when OPENAI_API_KEY is set and the embeddings store has entries.
- * Falls back silently - never blocks startup or errors out.
- */
 async function renderSemanticRecall(taskInput) {
   if (!taskInput || taskInput.length < 10) return null;
   try {
-    const { getRelevantMemories } = await import("./tools/memory.js");
+    const { getRelevantMemories } = await import("../tools/memory.js");
     const { tenantId } = _getContextMemoryPaths();
     return await getRelevantMemories(taskInput, 5, tenantId);
   } catch {
@@ -79,7 +101,7 @@ async function renderSemanticRecall(taskInput) {
   }
 }
 
-// --- Section Renderers ---
+// ── Section Renderers ────────────────────────────────────────────────────────
 
 function renderSoul() {
   if (existsSync(config.soulPath)) {
@@ -88,6 +110,31 @@ function renderSoul() {
   return "You are Daemora, a personal helpful AI assistant. Execute tasks immediately using tools.";
 }
 
+function renderUserProfile() {
+  const store = tenantContext.getStore();
+  const tenantId = store?.tenant?.id;
+  let profilePath;
+  if (tenantId) {
+    const safeId = tenantId.replace(/[^a-zA-Z0-9_-]/g, "_");
+    profilePath = join(config.dataDir, "tenants", safeId, "user-profile.json");
+  } else {
+    profilePath = join(config.dataDir, "user-profile.json");
+  }
+  if (!existsSync(profilePath)) return null;
+  try {
+    const profile = JSON.parse(readFileSync(profilePath, "utf-8"));
+    const lines = [];
+    if (profile.name) lines.push(`Name: ${profile.name}`);
+    if (profile.personality) lines.push(`Personality: ${profile.personality}`);
+    if (profile.tone) lines.push(`Tone: ${profile.tone}`);
+    if (profile.instructions) lines.push(`\nCustom Instructions:\n${profile.instructions}`);
+    if (lines.length === 0) return null;
+    return `# User Profile\n\n${lines.join("\n")}`;
+  } catch {
+    return null;
+  }
+}
+
 function renderResponseFormat() {
   return `# Response Format
 
@@ -129,6 +176,15 @@ You MUST respond with a JSON object matching this exact schema on every turn:
 }
 
 function renderToolDocs() {
+  const unconfigured = Object.keys(TOOL_REQUIRED_KEYS).filter(t => !_isToolConfigured(t));
+
+  // Build the "no auth" warning section for unconfigured tools
+  const noAuthSection = unconfigured.length > 0
+    ? `\n\n## Unconfigured Tools [NO AUTH]
+The following tools require API keys that are NOT set. **Do NOT call these tools.** If the user asks to use one, tell them to configure the required keys first (Settings page or \`daemora setup\`).
+${unconfigured.map(t => `- ${t} — needs: ${TOOL_REQUIRED_KEYS[t].join(" or ")}`).join("\n")}`
+    : "";
+
   return `# Available Tools
 
 All tool params are STRINGS. Pass them as an array of strings.
@@ -162,9 +218,9 @@ All tool params are STRINGS. Pass them as an array of strings.
   **Tabs**: newTab(url?), switchTab(targetId), listTabs, closeTab(targetId?).
   **Other**: resize(WxH), highlight(ref|selector), handleDialog(accept|dismiss,text?), newSession(profile?), status, close.
   Localhost/127.0.0.1 allowed. Use refs from snapshot instead of CSS selectors.
-
+${_isToolConfigured("sendEmail") ? `
 ## Communication
-- sendEmail(to, subject, body, optionsJson?) — Send email via SMTP. opts: {"cc":"...","bcc":"...","attachments":[...]}
+- sendEmail(to, subject, body, optionsJson?) — Send email via SMTP. opts: {"cc":"...","bcc":"...","attachments":[...]}` : ""}
 - messageChannel(channel, target, message) — Send message on any channel. channel: "telegram"|"whatsapp"|"email".
 
 ## Documents
@@ -173,8 +229,8 @@ All tool params are STRINGS. Pass them as an array of strings.
 ## Vision & Screen
 - imageAnalysis(imagePath, prompt?) — Analyze image with vision model. Path or URL.
 - screenCapture(optionsJson?) — Screenshot or video. opts: {"mode":"screenshot"|"video","outputDir":"/tmp","duration":10}. Chain with replyWithFile or imageAnalysis.
-- transcribeAudio(audioPath, prompt?) — Transcribe audio to text via Whisper. Formats: mp3, wav, m4a, webm, ogg, flac.
-- textToSpeech(text, optionsJson?) — Text to MP3. opts: {"voice":"nova|alloy|echo|fable|onyx|shimmer","provider":"openai|elevenlabs"}. Chain with replyWithFile.
+${_isToolConfigured("transcribeAudio") ? `- transcribeAudio(audioPath, prompt?) — Transcribe audio to text via Whisper. Formats: mp3, wav, m4a, webm, ogg, flac.` : ""}
+${_isToolConfigured("textToSpeech") ? `- textToSpeech(text, optionsJson?) — Text to MP3. opts: {"voice":"nova|alloy|echo|fable|onyx|shimmer","provider":"openai|elevenlabs"}. Chain with replyWithFile.` : ""}
 - replyWithFile(filePath, caption?) — Send file back to current user. Use for any generated file (screenshot, doc, audio).
 - sendFile(channel, target, filePath, caption?) — Send file to a DIFFERENT user on a specific channel.
 
@@ -188,7 +244,7 @@ All tool params are STRINGS. Pass them as an array of strings.
 - writeDailyLog(entry) — Append to today's daily log.
 
 ## Agents
-- spawnAgent(taskDescription, optionsJson?) — Spawn sub-agent. opts: {"profile":"coder|researcher|writer|analyst","extraTools":[...],"parentContext":"...","model":"..."}. Task description must be comprehensive — sub-agent has no other context.
+- spawnAgent(taskDescription, optionsJson?) — Spawn sub-agent. opts: {"profile":"coder|researcher|writer|analyst","extraTools":[...],"skills":["skills/coding.md"],"parentContext":"...","model":"..."}. Pass skills array with skill paths from the Available Skills list — the skill content is injected directly into the sub-agent so it can follow the instructions without loading them. Task description must be comprehensive — sub-agent has no other context.
 - parallelAgents(tasksJson, sharedOptionsJson?) — Spawn multiple agents in parallel. tasksJson: [{"description":"...","options":{...}}]. sharedOptionsJson: {"sharedContext":"..."}. Always pass workspace path in sharedContext.
 - manageAgents(action, paramsJson?) — List, kill, or steer agents. action: "list"|"kill"|"steer".
 
@@ -209,7 +265,7 @@ Delegate a task to a specialist agent for the named MCP server.
 - projectTracker(action, paramsJson?) — Track multi-step projects. Actions: createProject, addTask, updateTask, getProject, listProjects, deleteProject. Persisted to disk.
 
 ## Automation
-- cron(action, paramsJson?) — Schedule recurring tasks. action: "list"|"add"|"remove"|"run"|"status". opts for add: {"cronExpression":"...","taskInput":"...","name":"..."}`;
+- cron(action, paramsJson?) — Schedule recurring tasks. action: "list"|"add"|"remove"|"run"|"status". opts for add: {"cronExpression":"...","taskInput":"...","name":"..."}${noAuthSection}`;
 }
 
 function renderMCPTools() {
@@ -270,12 +326,11 @@ function renderToolUsageRules() {
 - Prefer simplest correct solution. Complexity is a cost.`;
 }
 
-async function renderSkills(taskInput) {
+async function renderSkills(taskInput, limit = 20) {
   const totalCount = skillLoader.list().length;
   if (totalCount === 0) return "";
 
-  // Hybrid: use embeddings/keyword matching to rank, show top 20
-  const summaries = await skillLoader.getMatchedSkillSummaries(taskInput, 20);
+  const summaries = await skillLoader.getMatchedSkillSummaries(taskInput, limit);
   if (!summaries || summaries.length === 0) return "";
 
   const lines = summaries.map(s =>
@@ -350,11 +405,14 @@ function renderSubagentContext(taskDescription) {
   if (!taskDescription) return null;
   return `# Subagent Context
 
-You are a sub-agent spawned for a specific task.
-- Complete your assigned task. That's your entire purpose.
-- Stay focused — no side quests, no proactive actions.
-- Your final message will be reported back to the parent agent.
-- Include: what you accomplished, relevant details, keep it concise.`;
+You are a sub-agent spawned for a specific task. Complete it fully without asking questions.
+
+## Rules
+- Execute the task end-to-end. Do not stop to ask the parent agent for clarification — figure it out.
+- If matched skills were injected in your context, follow them precisely.
+- If you need a skill not already injected, load it with \`readFile("skills/<name>.md")\` and follow its instructions.
+- Use every tool, command, and skill available to you to finish the job.
+- When done, report back: what you did, key outcomes, any issues found. Keep it concise.`;
 }
 
 function renderRuntime(meta = {}) {
@@ -366,6 +424,4 @@ function renderRuntime(meta = {}) {
   return `Runtime: ${parts.join(" | ")}`;
 }
 
-// Note: buildSystemPrompt is now async. Use `await buildSystemPrompt(taskInput)` at call sites.
-// This legacy sync export is kept for any import that doesn't need task-specific recall.
-export const systemPrompt = { role: "system", content: "" }; // placeholder - rebuilt per-task
+export const systemPrompt = { role: "system", content: "" };

package/src/cli.js

@@ -15,9 +15,11 @@ import chalk from "chalk";
 import { config } from "./config/default.js";
 import daemonManager from "./daemon/DaemonManager.js";
 import secretVault from "./safety/SecretVault.js";
-import { readFileSync, writeFileSync, existsSync } from "fs";
-import { join } from "path";
+import { readFileSync, writeFileSync, existsSync, mkdirSync } from "fs";
+import { join, dirname } from "path";
+import { fileURLToPath } from "url";
 import { execSync } from "child_process";
+import { randomBytes } from "crypto";
 
 // ── Color palette — matches Daemora UI exactly ──────────────────────────────
 const P = {
@@ -68,6 +70,14 @@ const [,, command, subcommand, ...rest] = process.argv;
 
 async function main() {
   switch (command) {
+    case "version":
+    case "--version":
+    case "-v": {
+      const pkg = JSON.parse(readFileSync(join(dirname(fileURLToPath(import.meta.url)), "..", "package.json"), "utf-8"));
+      console.log(`daemora v${pkg.version}`);
+      break;
+    }
+
     case "start":
       // If vault exists, prompt for passphrase and inject secrets before server boot
       if (secretVault.exists()) {
@@ -134,6 +144,14 @@ async function main() {
       await handleTools(subcommand);
       break;
 
+    case "config":
+      handleConfig(subcommand, rest);
+      break;
+
+    case "auth":
+      handleAuth(subcommand);
+      break;
+
     case "setup":
       const { runSetupWizard } = await import("./setup/wizard.js");
       await runSetupWizard();
@@ -425,7 +443,7 @@ async function handleMCP(action, args) {
           }));
           if (needsEnv) {
             serverConfig.env = {};
-            p.log.info(`  Tip: use \${MY_VAR} to reference existing env vars instead of pasting secrets`);
+            pi.log.info(`  Tip: use \${MY_VAR} to reference existing env vars instead of pasting secrets`);
             let more = true;
             while (more) {
               const key = pGuard(await pi.text({
@@ -649,6 +667,132 @@ async function handleMCP(action, args) {
   }
 }
 
+// ── Config (env var management from CLI) ──────────────────────────────────────
+
+function handleConfig(action, args) {
+  const header = `\n  ${t.h("Daemora Config")}  ${t.muted("Environment variable management")}\n`;
+
+  switch (action) {
+    case "set": {
+      const [key, ...valueParts] = args;
+      const value = valueParts.join(" ");
+      if (!key || !value) {
+        console.error(`\n  ${S.cross}  Usage: daemora config set ${t.dim("<KEY> <value>")}\n`);
+        console.log(`  ${t.muted("Example:")}  daemora config set OPENAI_API_KEY sk-...\n`);
+        process.exit(1);
+      }
+      writeEnvKey(key, value);
+      process.env[key] = value;
+      console.log(`${header}  ${S.check}  ${t.success(key)} = ${t.muted(value.length <= 8 ? value : value.slice(0, 4) + "****")}\n`);
+      break;
+    }
+    case "get": {
+      const [key] = args;
+      if (!key) {
+        console.error(`\n  ${S.cross}  Usage: daemora config get ${t.dim("<KEY>")}\n`);
+        process.exit(1);
+      }
+      const env = readEnvFile();
+      const val = env[key];
+      if (val !== undefined) {
+        const masked = val.length <= 4 ? "****" : val.slice(0, 4) + "*".repeat(Math.min(val.length - 4, 20));
+        console.log(`${header}  ${key} = ${t.muted(masked)}\n`);
+      } else {
+        console.log(`${header}  ${S.cross}  ${key} is not set\n`);
+      }
+      break;
+    }
+    case "delete":
+    case "unset": {
+      const [key] = args;
+      if (!key) {
+        console.error(`\n  ${S.cross}  Usage: daemora config unset ${t.dim("<KEY>")}\n`);
+        process.exit(1);
+      }
+      deleteEnvKey(key);
+      delete process.env[key];
+      console.log(`${header}  ${S.check}  ${key} removed\n`);
+      break;
+    }
+    case "list":
+    default: {
+      const env = readEnvFile();
+      const keys = Object.keys(env);
+      console.log(header);
+      if (keys.length === 0) {
+        console.log(`  ${t.muted("No env vars configured. Run:")}  daemora config set <KEY> <value>\n`);
+      } else {
+        // Also read .env.example for available keys
+        const examplePath = join(config.rootDir, ".env.example");
+        const availableKeys = new Set();
+        if (existsSync(examplePath)) {
+          for (const line of readFileSync(examplePath, "utf-8").split("\n")) {
+            const trimmed = line.trim();
+            if (!trimmed || trimmed.startsWith("#")) continue;
+            const eqIdx = trimmed.indexOf("=");
+            if (eqIdx > 0) availableKeys.add(trimmed.slice(0, eqIdx));
+          }
+        }
+
+        // Show configured keys
+        console.log(`  ${t.muted("Configured")}  (${keys.length} keys)\n`);
+        for (const key of keys) {
+          const val = env[key];
+          const masked = !val ? t.dim("(empty)") : val.length <= 4 ? "****" : val.slice(0, 4) + "*".repeat(Math.min(val.length - 4, 16));
+          console.log(`  ${S.check}  ${t.success(key.padEnd(30))} ${t.muted(masked)}`);
+        }
+
+        // Show unconfigured keys from .env.example
+        const unconfigured = [...availableKeys].filter(k => !env[k]);
+        if (unconfigured.length > 0) {
+          console.log(`\n  ${t.muted("Available (not set)")}  (${unconfigured.length} keys)\n`);
+          for (const key of unconfigured.slice(0, 20)) {
+            console.log(`  ${S.cross}  ${t.dim(key)}`);
+          }
+          if (unconfigured.length > 20) {
+            console.log(`  ${t.dim(`... and ${unconfigured.length - 20} more`)}`);
+          }
+        }
+        console.log("");
+      }
+      break;
+    }
+  }
+}
+
+// ── Auth (API token management) ───────────────────────────────────────────────
+
+function handleAuth(action) {
+  const tokenPath = join(config.dataDir, "auth-token");
+  const header = `\n  ${t.h("Daemora Auth")}  ${t.muted("API token management")}\n`;
+
+  switch (action) {
+    case "token": {
+      if (!existsSync(tokenPath)) {
+        console.log(`${header}  ${S.cross}  No token yet. Start the server first or run: daemora auth reset\n`);
+      } else {
+        const token = readFileSync(tokenPath, "utf-8").trim();
+        console.log(`${header}  ${t.muted("API Token:")}\n\n  ${token}\n`);
+        console.log(`  ${t.muted("Usage:")}  curl -H "Authorization: Bearer ${token.slice(0, 8)}..." http://localhost:${config.port}/api/health\n`);
+      }
+      break;
+    }
+    case "reset": {
+      const token = randomBytes(32).toString("hex");
+      mkdirSync(dirname(tokenPath), { recursive: true });
+      writeFileSync(tokenPath, token, { mode: 0o600 });
+      console.log(`${header}  ${S.check}  ${t.success("New token generated")}\n\n  ${token}\n`);
+      console.log(`  ${t.muted("Restart the server for the new token to take effect.")}\n`);
+      break;
+    }
+    default: {
+      console.log(`${header}  ${t.cmd("daemora auth token")}    Show current API token`);
+      console.log(`  ${t.cmd("daemora auth reset")}    Generate a new token\n`);
+      break;
+    }
+  }
+}
+
 // ── Sandbox (filesystem scoping) helpers ──────────────────────────────────────
 
 function readEnvFile() {
@@ -2060,6 +2204,15 @@ ${line}
   ${t.cmd("start")}                            Start the agent server
   ${t.cmd("setup")}                            Interactive setup wizard
 
+  ${t.cmd("auth token")}                        Show API auth token
+  ${t.cmd("auth reset")}                        Generate a new auth token
+
+  ${t.cmd("config list")}                      List all configured env vars
+  ${t.cmd("config set")} ${t.dim("<KEY> <value>")}          Set an env var (e.g. OPENAI_API_KEY)
+  ${t.cmd("config get")} ${t.dim("<KEY>")}                  Show an env var (masked)
+  ${t.cmd("config unset")} ${t.dim("<KEY>")}                Remove an env var
+  ${t.dim("  Keys: DEFAULT_MODEL, SUB_AGENT_MODEL, CODE_MODEL, RESEARCH_MODEL ...")}
+
   ${t.cmd("daemon install")}                   Install as OS service (auto-start)
   ${t.cmd("daemon uninstall")}                 Remove OS service
   ${t.cmd("daemon start")}                     Start the background daemon
@@ -2116,6 +2269,7 @@ ${line}
   ${t.cmd("cleanup set")} ${t.dim("<days>")}             Set auto-cleanup retention (0 = never)
   ${t.cmd("cleanup stats")}                    Show storage usage per directory
 
+  ${t.cmd("version")}  ${t.dim("-v --version")}             Show version
   ${t.cmd("help")}                             Show this help
 
   ${t.bold("EXAMPLES")}
@@ -2132,8 +2286,11 @@ ${line}
   ${t.dim("$")} daemora mcp list
   ${t.dim("$")} daemora mcp add github npx -y @modelcontextprotocol/server-github
   ${t.dim("$")} daemora mcp env github GITHUB_PERSONAL_ACCESS_TOKEN ghp_...
-  ${t.dim("$")} daemora mcp add notion http://localhost:3100/mcp
-  ${t.dim("$")} daemora mcp add myserver http://localhost:3100/sse --sse
+  ${t.dim("$")} daemora mcp env notion NOTION_TOKEN ntn_...
+  ${t.dim("$")} daemora mcp env stripe STRIPE_SECRET_KEY sk_live_...
+  ${t.dim("$")} daemora mcp enable notion
+  ${t.dim("$")} daemora mcp add myserver https://api.example.com/mcp
+  ${t.dim("$")} daemora mcp add mysse https://api.example.com/sse --sse
   ${t.dim("$")} daemora mcp remove github
   ${t.dim("$")} daemora mcp add                   (interactive - prompts for everything)
   ${t.dim("$")} daemora mcp reload github         (reconnects live if agent running)

package/src/config/default.js

@@ -28,7 +28,11 @@ export const config = {
   memoryPath: join(ROOT_DIR, "MEMORY.md"),
 
   // Default model (provider:model format)
-  defaultModel: process.env.DEFAULT_MODEL || "openai:gpt-4.1-mini",
+  defaultModel: process.env.DEFAULT_MODEL || "openai:gpt-5.1-mini",
+
+  // Sub-agent model — used for all sub-agents when no profile-specific model is set.
+  // Falls between profile routing (CODE_MODEL etc.) and DEFAULT_MODEL in priority.
+  subAgentModel: process.env.SUB_AGENT_MODEL || null,
 
   // Agent loop
   maxLoops: 40,

package/src/core/Compaction.js

@@ -145,14 +145,18 @@ Available tools: ${Object.keys(memoryTools).join(", ")}`;
 export async function compactIfNeeded(messages, modelMeta, taskId = null, tools = {}) {
   const tokenCount = estimateTokens(messages);
 
-  if (tokenCount < modelMeta.compactAt) {
+  // Dynamic threshold: compact when within 10k tokens of the model's context window
+  const contextLimit = modelMeta.contextWindow || 128_000;
+  const compactThreshold = Math.max(contextLimit - 10_000, modelMeta.compactAt || 90_000);
+
+  if (tokenCount < compactThreshold) {
     return messages;
   }
 
   console.log(
-    `[Compaction] Triggered: ~${tokenCount} tokens exceeds threshold ${modelMeta.compactAt}`
+    `[Compaction] Triggered: ~${tokenCount} tokens exceeds threshold ${compactThreshold} (context: ${contextLimit})`
   );
-  eventBus.emitEvent("compact:triggered", { tokenCount, threshold: modelMeta.compactAt });
+  eventBus.emitEvent("compact:triggered", { tokenCount, threshold: compactThreshold });
 
   // Pre-compaction memory flush — let agent save important context before we compact
   await runPreCompactionFlush(messages, tools);
@@ -182,15 +186,29 @@ export async function compactIfNeeded(messages, modelMeta, taskId = null, tools
     return msg;
   });
 
-  // Step 3: Summarize old messages using a cheap model
+  // Step 3: Summarize old messages using the same model (or cheap fallback)
   try {
-    const { model } = getCheapModel();
-    const summaryPrompt = `Summarize the following conversation history concisely. Preserve:
-- Key decisions made
+    let model;
+    try {
+      // Prefer the same model the agent is using for consistent quality
+      const { getModelWithFallback } = await import("../models/ModelRouter.js");
+      const resolved = getModelWithFallback(modelMeta.provider ? `${modelMeta.provider}:${modelMeta.model}` : null);
+      model = resolved.model;
+    } catch {
+      const cheap = getCheapModel();
+      model = cheap.model;
+    }
+
+    const summaryPrompt = `Summarize the following conversation history concisely. You MUST preserve:
+- What was done (completed steps)
+- What is left to do (pending work, next steps)
+- Key decisions made and why
 - File paths mentioned and their purpose
-- Task progress and what was accomplished
 - Any errors encountered and how they were resolved
-- User preferences or instructions
+- User preferences, instructions, or constraints
+- Current task status (in progress, blocked, etc.)
+
+Format as a structured summary with clear sections.
 
 Conversation to summarize:
 ${prunedOld.map((m) => `[${m.role}]: ${typeof m.content === "string" ? m.content.slice(0, 2000) : JSON.stringify(m.content).slice(0, 2000)}`).join("\n")}`;

package/src/core/TaskRunner.js

@@ -1,5 +1,5 @@
 import { runAgentLoop } from "./AgentLoop.js";
-import { buildSystemPrompt } from "../systemPrompt.js";
+import { buildSystemPrompt } from "../agents/systemPrompt.js";
 import { toolFunctions } from "../tools/index.js";
 import { createSession, getSession, setMessages } from "../services/sessions.js";
 import taskQueue from "./TaskQueue.js";