titan-agent 7.0.0 → 7.2.0

package/README.md

@@ -14,7 +14,7 @@
 </p>
 
 <p align="center">
-  <strong>The local-first AI agent framework that is model-agnostic, accessible by design, and lets you watch it work.</strong>
+  <strong>The local-first AI agent framework that teaches itself, mixes councils of models, plays well with other agents — and lets you watch it work.</strong>
   <br><small>Your hardware. Your models. Your control. Built in TypeScript, MIT-licensed, 40K+ lifetime installs.</small>
 </p>
 
@@ -50,6 +50,56 @@
 
 <a id="whats-new"></a>
 
+## What's new in v7.2.0 "Conscience" — Honesty as an Enforced Invariant
+
+**TITAN stopped being able to bluff.** Two honesty organs turn any model — local
+or cloud — into one that *verifies before it claims* and *critiques before it
+speaks*. Flip one switch (`agent.reliabilityMode: true`) and your local model
+behaves with the reliability discipline of a frontier model, because reliability
+is enforced in the harness, not hoped for in the prompt.
+
+- 🛡️ **Verification wall (always on):** claim you sent/posted/deleted/deployed/
+  saved something with no tool that did it → TITAN appends a visible correction.
+- 🧠 **Self-critique (Reliability Mode):** the model reviews its own draft
+  adversarially and appends honest "on reflection" caveats. Live-proven: a local
+  model caught its own over-precise number and a definitional conflation,
+  unprompted.
+- ⚡ **Perf:** ~19MB/turn read deleted (tail-read trajectories); every timer
+  `.unref()`'d.
+
+See **[CHANGELOG.md](CHANGELOG.md)** for the full v7.2.0 entry.
+
+---
+
+## What's new in v7.1.0 "Council" — Local-First Intelligence, Together
+
+**Agents advising agents, models sized to their machines, and benchmarks you can trust.**
+
+### 🧠 `/moa` — Mixture of Agents
+Ask a **council**, not a model: N reference advisors answer in parallel (trimmed context, no tools, per-advisor timeouts), and one acting aggregator synthesizes the answer with full tool use. Advisors default to **your own local models across your own machines** — the mixture costs nothing and parallelizes across GPUs. `/moa <prompt>` for one answer, `/moa use <preset>` for a session, presets in config.
+
+### 🌐 Works alongside Hermes Agent & OpenClaw — the MCP triangle
+TITAN's MCP server now speaks **whole-agent**: `titan_chat`, `titan_delegate_task` / `titan_task_status`, `titan_moa`, `titan_status`. One config line makes TITAN a tool inside Hermes or OpenClaw — and TITAN's MCP client consumes their servers right back. Six directed edges, all shipped surfaces, no adapters. ([docs/INTEROP.md](docs/INTEROP.md))
+
+### 📏 Context-Fit — local models become first-class
+TITAN now **learns each deployment's real context ceiling from live traffic** and sizes its toolset to fit: small-context deployments get a lean toolset instead of a fatal overflow, and every request got ~4K tokens lighter (the tool catalog shrank 5×). Strict local backends (vLLM/llama.cpp/LM Studio template quirks) are handled; gateway misroutes fail fast.
+
+**Proof, not vibes**: under the new `TITAN_BENCH=1` isolation mode, **Qwen3.6-35B-A3B NVFP4 on DGX-Spark-class hardware scores 85% through TITAN's harness — tying the RTX 5090's best local model.** Full corrected tables in [benchmarks/MODEL_COMPARISON.md](benchmarks/MODEL_COMPARISON.md).
+
+### 🛡️ Honesty as an enforced invariant (any model)
+TITAN won't let a model bluff. The **verification wall** appends a visible
+correction if a reply claims it *did* something (sent, posted, deleted,
+deployed, saved) with no tool that actually did it — deterministic, always on.
+Turn on **Reliability Mode** (`agent.reliabilityMode: true` — one switch) and TITAN
+also critiques its own draft before you see it, appending honest "on reflection"
+caveats for anything it didn't verify. This is how a local model reaches
+frontier *reliability* without frontier size — and it works on whatever brain
+you plug in.
+
+See **[CHANGELOG.md](CHANGELOG.md)** for the complete v7.1.0 entry.
+
+---
+
 ## What's new in v7.0.0 "Independence" — Alive, Welcoming, Model-Agnostic
 
 **The Independence Day release** — because local-first *is* independence: your AI, your hardware, your data. Four themes: **TITAN provably teaches itself from your usage, first run takes one minute, it lives on your machine between conversations, and it runs well on any capable model.** (Connect your first model and you'll see the fireworks. 🎆)
@@ -148,6 +198,8 @@ See **[CHANGELOG.md](CHANGELOG.md)** for the full v7.0.0 entry and everything be
 ## Why TITAN
 
 - **It teaches itself — provably.** Muscle Memory turns your repeated workflows into replay-verified skills. Every learned skill passes a deterministic exam against your real usage before you see it; nothing auto-adopts. No other framework ships trustworthy self-improvement on by default.
+- **A council, not just a model.** `/moa` fans your question to advisor models across your own machines and synthesizes one sharper answer — mixture-of-agents at $0/turn.
+- **Plays well with other agents.** TITAN is an MCP tool inside Hermes/OpenClaw and consumes them right back — delegate jobs, share councils, cooperate.
 - **One-minute first run.** `npm install -g titan-agent && titan gateway` — the dashboard greets you and connects a model in about a minute. No Docker, no YAML, no config-file editing.
 - **Model-agnostic, for real.** One harness, native tool-calls per vendor, a per-model capability registry, adaptive token ceilings. Bring DeepSeek, Qwen, GLM, Kimi, MiniMax, Anthropic, OpenAI, Gemini, or anything OpenAI-compatible — the framework adapts instead of being tuned to one LLM.
 - **Local-first.** Run any model on your own hardware via Ollama, or route to any of 36 providers through one router. Switch models mid-conversation. Failed models auto-fallback. Your data never leaves your machine unless you explicitly send it to a cloud provider.
@@ -168,7 +220,7 @@ titan gateway            # opens Mission Control at http://localhost:48420
 
 That's it — no setup required first. The dashboard greets you and connects a model in about a minute (one click if Ollama is running; or paste any API key / OpenAI-compatible endpoint). Prefer the terminal? `titan onboard` runs the full interactive wizard.
 
-**Which model?** Benchmarked through TITAN's own harness (July 2026): best local = **`qwen3-coder-next`**-class (74% harness pass, fastest median case of anything tested); best cloud = **GLM-5.1** / **Kimi K2.6** (93%, both aced safety). Mid-size models on 32k contexts overflow TITAN's full toolset — details and honest caveats in [benchmarks/MODEL_COMPARISON.md](benchmarks/MODEL_COMPARISON.md).
+**Which model?** Benchmarked through TITAN's own harness (July 2026, clean rerun): best local = **`qwen3-coder-next`** (85%, 5.2s median) and — on DGX-Spark-class unified memory — **Qwen3.6-35B-A3B NVFP4** (85%, ties it); best cloud = **GLM-5.1** / **Kimi K2.6** (93%). TITAN v7.1 auto-adapts to each deployment's real context (learned live), so small-context deployments degrade gracefully instead of failing — details and honest caveats in [benchmarks/MODEL_COMPARISON.md](benchmarks/MODEL_COMPARISON.md).
 
 Or one-liner:
 

package/dist/agent/agent.js

@@ -1294,12 +1294,43 @@ CRITICAL: Widgets are NOT files. Do NOT use write_file, edit_file, or any file t
       }
     }
     const SMALL_MODEL_PATTERNS = ["llama3.2", "llama3.1:8b", "phi", "gemma:2b", "qwen3.5:4b", "tinyllama", "dolphin3"];
-    const isSmallModel = SMALL_MODEL_PATTERNS.some((p) => activeModel.toLowerCase().includes(p));
-    if (isSmallModel) {
+    const nameIsTiny = SMALL_MODEL_PATTERNS.some((p) => activeModel.toLowerCase().includes(p));
+    const { getModelCapabilitiesEx } = await import("../providers/modelCapabilities.js");
+    const bareModel = activeModel.includes("/") ? activeModel.slice(activeModel.indexOf("/") + 1) : activeModel;
+    const capsA = getModelCapabilitiesEx(activeModel);
+    const capsB = getModelCapabilitiesEx(bareModel);
+    const best = capsA.caps.contextWindow <= capsB.caps.contextWindow ? capsA : capsB;
+    const modelCtxWindow = best.caps.contextWindow;
+    const ctxIsKnown = best.source !== "fallback";
+    const contextTier = nameIsTiny || ctxIsKnown && modelCtxWindow < 12e3 ? "tiny" : ctxIsKnown && modelCtxWindow < 48e3 ? "small" : "full";
+    const isSmallModel = contextTier !== "full";
+    if (contextTier === "tiny") {
       const CORE_TOOL_NAMES = ["shell", "read_file", "write_file", "edit_file", "list_dir", "memory"];
       const coreTools = activeTools.filter((t) => CORE_TOOL_NAMES.includes(t.function.name));
-      logger.info(COMPONENT, `[SmallModel] Reducing tools from ${activeTools.length} to ${coreTools.length} for ${activeModel}`);
+      logger.info(COMPONENT, `[ContextFit] tier=tiny (ctx=${modelCtxWindow}) \u2014 ${activeTools.length} \u2192 ${coreTools.length} tools for ${activeModel}`);
       activeTools = coreTools;
+    } else if (contextTier === "small") {
+      const LEAN_TOOL_NAMES = [
+        "shell",
+        "read_file",
+        "write_file",
+        "edit_file",
+        "list_dir",
+        "web_search",
+        "browse_url",
+        "memory",
+        "create_widget",
+        "gallery_search",
+        "gallery_get",
+        "cron",
+        "reminder",
+        "tool_search"
+      ];
+      const leanTools = activeTools.filter((t) => LEAN_TOOL_NAMES.includes(t.function.name));
+      if (leanTools.length > 0) {
+        logger.info(COMPONENT, `[ContextFit] tier=small (ctx=${modelCtxWindow}) \u2014 ${activeTools.length} \u2192 ${leanTools.length} tools for ${activeModel}`);
+        activeTools = leanTools;
+      }
     }
     if (!voiceFastPath && !isSmallModel && isBrainAvailable()) {
       const brainFiltered = await brainSelectTools(message, activeTools);
@@ -1511,12 +1542,23 @@ ${buildSystemWidgetGate(sw)}` : buildSystemWidgetGate(sw);
       logger.info(COMPONENT, `[SystemWidgetShortcut] Deterministically emitted _____widget gate for ${sw.source}`);
     }
   }
+  try {
+    const agentCfg = config.agent;
+    const { resolveSelfCritique, runSelfCritique } = await import("./selfCritique.js");
+    const scCfg = resolveSelfCritique(agentCfg.selfCritique, agentCfg.reliabilityMode === true);
+    if (scCfg.enabled) {
+      const sc = await runSelfCritique(scCfg, message, finalContent, [...new Set(toolsUsed)], modelUsed);
+      if (sc.critiqued && sc.issues.length > 0) finalContent = sc.content;
+    }
+  } catch (e) {
+    logger.debug(COMPONENT, `Self-critique wrapper skipped: ${e.message}`);
+  }
   {
-    const claimsScheduled = /\b(i(?:'|’)?(?:ve| have)\s+(?:set|scheduled|created)|(?:reminder|schedule[d]?)\s+(?:is\s+)?(?:set|created|scheduled)|i(?:'|’)?ll remind you|scheduled a reminder)\b/i.test(finalContent);
-    const schedulingToolRan = toolsUsed.some((t) => /^(reminder|cron|schedule|event_trigger|workflows?|social_scheduler)/i.test(t));
-    if (claimsScheduled && !schedulingToolRan) {
-      finalContent += "\n\n\u26A0\uFE0F Correction: I described scheduling that, but I did not actually create a scheduled job this turn. Ask me again and I'll set it with the reminder tool properly.";
-      logger.warn(COMPONENT, "[HonestyGuard] Reply claimed scheduling without a scheduling tool \u2014 appended correction");
+    const { applyHonestyGuard } = await import("./honestyGuard.js");
+    const guarded = applyHonestyGuard(finalContent, toolsUsed);
+    if (guarded.flagged.length > 0) {
+      finalContent = guarded.content;
+      logger.warn(COMPONENT, `[HonestyGuard] Reply claimed ${guarded.flagged.join(", ")} without a capable tool \u2014 appended correction`);
     }
   }
   addMessage(session, "assistant", finalContent, {
@@ -1647,8 +1689,11 @@ ${buildSystemWidgetGate(sw)}` : buildSystemWidgetGate(sw);
   } catch {
   }
   try {
-    const { maybeObserve } = await import("./observationWriter.js");
-    maybeObserve(channel, message || "", finalContent || "");
+    const { isBenchMode } = await import("../utils/benchMode.js");
+    if (!isBenchMode()) {
+      const { maybeObserve } = await import("./observationWriter.js");
+      maybeObserve(channel, message || "", finalContent || "");
+    }
   } catch {
   }
   return {