chore(release): 1.25.0 [skip ci]

# [1.25.0](https://github.com/asepharyana/zesdex/compare/v1.24.0...v1.25.0) (2026-09-03)

### Features

* **tui:** enhance LLM interaction with max_tokens and temperature overrides; add tests for parallel delegation and workflow script parsing ([5c65f8f](https://github.com/asepharyana/zesdex/commit/5c65f8f6ab093dbfad682939a9fcae79a8a68661))

CHANGELOG.md

@@ -1,3 +1,30 @@
+# [1.25.0](https://github.com/asepharyana/zesdex/compare/v1.24.0...v1.25.0) (2026-09-03)
+
+
+### Features
+
+* **tui:** enhance LLM interaction with max_tokens and temperature overrides; add tests for parallel delegation and workflow script parsing ([5c65f8f](https://github.com/asepharyana/zesdex/commit/5c65f8f6ab093dbfad682939a9fcae79a8a68661))
+
+# [1.24.0](https://github.com/asepharyana/zesdex/compare/v1.23.2...v1.24.0) (2026-09-03)
+
+
+### Features
+
+* **tui:** add new flow helpers for bash exit code handling and verification prompts ([f98030f](https://github.com/asepharyana/zesdex/commit/f98030f28a20e3d9ba807bb98ddbf142511a3964))
+* **tui:** enhance autocomplete functionality and UI interactions ([008f737](https://github.com/asepharyana/zesdex/commit/008f737e1f5d0eef6a07ebe0fff2c9730bd44eb4))
+* **tui:** enhance command execution on autocomplete and improve UI responsiveness ([f5b0421](https://github.com/asepharyana/zesdex/commit/f5b04214cf385ba2e4a02f6ce202436ffab23c35))
+* **tui:** enhance model selection and improve UI components ([c53fdfa](https://github.com/asepharyana/zesdex/commit/c53fdfa7c99887e8f3dfe68e7f1b2e766de039b1))
+* **tui:** implement interactive menu system for model and config selection ([2100e7a](https://github.com/asepharyana/zesdex/commit/2100e7ab54460f28b67c3bb07855eef8f04eb9ec))
+* **tui:** implement live model fetching and enhance model selection menus ([93e5464](https://github.com/asepharyana/zesdex/commit/93e54643703ce7f18286bd6a988e0fab18cc9e66))
+* **tui:** implement message compaction and summarization logic for efficient conversation handling ([e6718fa](https://github.com/asepharyana/zesdex/commit/e6718fa11825cccc74a9d100755226b7dfefc106))
+
+## [1.23.2](https://github.com/asepharyana/zesdex/compare/v1.23.1...v1.23.2) (2026-09-03)
+
+
+### Bug Fixes
+
+* **tui:** sliding-window stream reconciler + UI polish ([d86d6ae](https://github.com/asepharyana/zesdex/commit/d86d6aeef2dcd4924ced14f21502a84f7e916a26))
+
 ## [1.23.1](https://github.com/asepharyana/zesdex/compare/v1.23.0...v1.23.1) (2026-09-03)
 
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "zesdex",
-  "version": "1.23.1",
+  "version": "1.25.0",
   "private": true,
   "type": "module",
   "packageManager": "bun@1.3.14",

src/features/agent/application/turn_service.test.ts

@@ -4,8 +4,25 @@ import {
   conversationChars,
   ErrorTracker,
   truncateToolOutput,
+  bashExitCode,
+  isToolFailure,
+  deriveVerifyCommand,
+  AgentTurnServiceImpl,
+  takeRecentTail,
+  reduceMessagesToDigest,
+  compactMessagesWithAi,
 } from "./turn_service.ts";
-import { type ChatMessage, newConversation, systemMessage, userMessage, toolResultMessage } from "@zesdex/domain";
+import {
+  type ChatMessage,
+  newConversation,
+  systemMessage,
+  userMessage,
+  assistantMessage,
+  toolResultMessage,
+} from "@zesdex/domain";
+import type { AgentTurnParams } from "@zesdex/agent";
+import type { ToolExecutor } from "./index.ts";
+import type { ProviderService } from "./ports.ts";
 
 describe("truncateToolOutput", () => {
   it("short output is unchanged", () => {
@@ -56,4 +73,314 @@ describe("conversationChars", () => {
     conv.messages.push(systemMessage("sys"), userMessage("hello world"), toolResultMessage("id", "output"));
     expect(conversationChars(conv.messages)).toBe(3 + 11 + 6);
   });
-});
+});
+/* ── New flow helpers ─────────────────────────────────────────────── */
+
+describe("bashExitCode / isToolFailure", () => {
+  it("extracts a non-zero exit code", () => {
+    expect(bashExitCode("boom\n\nExit code: 1 (1s)")).toBe(1);
+    expect(bashExitCode("done\n\nExit code: 0 (0.5s)")).toBe(0);
+  });
+  it("returns null when no exit-code line exists", () => {
+    expect(bashExitCode("plain output")).toBeNull();
+  });
+  it("isToolFailure flags bash non-zero exits as failures", () => {
+    expect(isToolFailure("bash", "nope\n\nExit code: 2 (1s)")).toBe(true);
+    expect(isToolFailure("bash", "ok\n\nExit code: 0 (1s)")).toBe(false);
+  });
+  it("isToolFailure still flags Error: prefixes for other tools", () => {
+    expect(isToolFailure("read", "Error: no such file")).toBe(true);
+  });
+});
+
+describe("deriveVerifyCommand", () => {
+  it("defaults to bun check+test for a Bun repo", () => {
+    // No real manifests in a control dir we guarantee to not exist.
+    expect(deriveVerifyCommand("/nonexistent-zesdex-dir")).toContain("bun");
+  });
+});
+
+/* ── runTurn flow tests (fakes, no network) ───────────────────────── */
+
+interface ScriptedStep {
+  content?: string | null;
+  tools?: Array<{ name: string; args?: string; id?: string }>;
+}
+
+/** A ProviderService that replays scripted chatStream responses. */
+function makeScriptedProvider(script: ScriptedStep[]): {
+  provider: ProviderService;
+  seen: Array<ChatMessage[]>;
+  chatCalls: Array<ChatMessage[]>;
+} {
+  const seen: Array<ChatMessage[]> = [];
+  const chatCalls: Array<ChatMessage[]> = [];
+  let step = 0;
+  const chatResponses: string[] = [];
+  const provider: ProviderService = {
+    async chat(messages, _tools, _maxTokens, _temperature) {
+      // Used by compaction and review. For review tests we script below.
+      chatCalls.push([...messages]);
+      const text = chatResponses.shift() ?? "";
+      return { message: { role: "assistant", content: text || null }, usage: [10, 5] };
+    },
+    async chatStream(messages, _tools, _max, _temp, onEvent) {
+      // Snapshot a copy: runTurn freely mutates the live array (splice/compact).
+      seen.push([...messages]);
+      const s = script[Math.min(step, script.length - 1)] ?? { content: "done" };
+      step += 1;
+      if (s.content !== undefined && s.content !== null) onEvent({ kind: "token", content: s.content });
+      if (s.tools && s.tools.length > 0) {
+        const msg = assistantMessage(null) as ChatMessage;
+        msg.tool_calls = s.tools.map((t, i) => ({
+          id: t.id ?? `call-${i}`,
+          type: "function",
+          function: { name: t.name, arguments: t.args ?? "{}" },
+        }));
+        return { message: msg, usage: [10, 2] };
+      }
+      return {
+        message: { role: "assistant", content: s.content ?? null },
+        usage: [10, 2],
+      };
+    },
+    // expose a way for tests to script chat replies
+  } as ProviderService;
+  (provider as unknown as { setChatReply: (s: string) => void }).setChatReply = (text: string) => {
+    chatResponses.push(text);
+  };
+  return { provider, seen, chatCalls };
+}
+
+/** A ToolExecutor that echoes fixed outputs per tool. */
+function fixedExecutor(outs: Record<string, string>): ToolExecutor {
+  return {
+    async execute(name) {
+      return outs[name] ?? "ok";
+    },
+    isParallelSafe() {
+      return false;
+    },
+  };
+}
+
+function buildParams(userText: string): {
+  params: AgentTurnParams;
+  events: Array<any>;
+} {
+  const events: Array<any> = [];
+  const params: AgentTurnParams = {
+    messages: [userMessage(userText)],
+    session_dir: "/tmp/zesdex-flow-test",
+    workspace_roots: ["/nonexistent-zesdex-dir"], // no AGENTS.md/package.json side effects
+    turn_events: { push: (e) => events.push(e), drain: () => [] },
+    in_flight: { value: false },
+    abort: new AbortController(),
+    api_key: "k",
+    model: "m",
+    api_base: "https://x",
+  };
+  return { params, events };
+}
+
+describe("runTurn", () => {
+  it("runs a simple one-shot prompt to completion", async () => {
+    const { provider } = makeScriptedProvider([{ content: "here is the answer" }]);
+    const svc = new AgentTurnServiceImpl(provider as never, fixedExecutor({}), [] as never);
+    const { params } = buildParams("hello");
+    await svc.runTurn(params);
+    expect(params.in_flight.value).toBe(false);
+  });
+});
+
+describe("runTurn verify-after-edit", () => {
+  it("injects a verify nudge after a write, cleared after a successful bash", async () => {
+    const { provider, seen } = makeScriptedProvider([
+      { tools: [{ name: "write" }] },
+      { tools: [{ name: "bash" }] },
+      { content: "done" },
+    ]);
+    const executor = fixedExecutor({ write: "wrote it", bash: "ok\n\nExit code: 0 (1s)" });
+    const svc = new AgentTurnServiceImpl(provider as never, executor, [] as never);
+    const { params } = buildParams("add a comment");
+    await svc.runTurn(params);
+    // The LLM call after the write should carry the verify nudge.
+    const second = seen[1] ?? [];
+    expect(second.some((m) => (m.content ?? "").includes("Run the verify command"))).toBe(true);
+    // The successful bash exits 0, which clears pendingVerify — so the nudge is
+    // injected exactly once, not repeated on every later call.
+    const finalBatch = seen[seen.length - 1] ?? [];
+    const nudgeCount = finalBatch.filter((m) => (m.content ?? "").includes("Run the verify command")).length;
+    expect(nudgeCount).toBe(1);
+  });
+});
+
+describe("runTurn convergence guard", () => {
+  it("stops after repeated identical read calls", async () => {
+    const script: ScriptedStep[] = [];
+    for (let i = 0; i < 6; i++) script.push({ tools: [{ name: "read" }] });
+    const { provider, seen } = makeScriptedProvider(script);
+    const executor = fixedExecutor({ read: "same content" });
+    const svc = new AgentTurnServiceImpl(provider as never, executor, [] as never);
+    const { params, events } = buildParams("read something");
+    await svc.runTurn(params);
+    const warned = events.some((e) => e.kind === "system_note" && /without any progress|repeated/.test(e.message ?? ""));
+    expect(warned).toBe(true);
+    // Not every call got issued — the guard broke early.
+    expect(seen.length).toBeLessThan(script.length);
+  });
+});
+
+describe("runTurn self-review", () => {
+  it("feeds a reviewer critique back as a system message after a mutator", async () => {
+    const { provider, seen } = makeScriptedProvider([
+      { tools: [{ name: "edit" }] },
+      { content: "fixed" },
+    ]);
+    const withChat = provider as unknown as { setChatReply(s: string): void };
+    withChat.setChatReply("- [PRIORITY: high] handle empty input in parse()");
+    const executor = fixedExecutor({ edit: "edited" });
+    const svc = new AgentTurnServiceImpl(provider as never, executor, [] as never);
+    const { params, events } = buildParams("fix parse()");
+    await svc.runTurn(params);
+    // The reviewer critique must have reached the next LLM call's history.
+    const second = seen[1] ?? [];
+    expect(second.some((m) => (m.content ?? "").includes("[Reviewer]"))).toBe(true);
+    // review_usage emitted.
+    expect(events.some((e) => e.kind === "review_usage")).toBe(true);
+  });
+
+  it("passes the work-in-progress to the reviewer (not an empty slate)", async () => {
+    const { provider, chatCalls } = makeScriptedProvider([
+      { tools: [{ name: "edit" }] },
+      { content: "fixed" },
+    ]);
+    (provider as unknown as { setChatReply(s: string): void }).setChatReply(
+      "- [PRIORITY: medium] add a null check",
+    );
+    const svc = new AgentTurnServiceImpl(provider as never, fixedExecutor({ edit: "changed" }), [] as never);
+    const { params } = buildParams("harden the parser");
+    await svc.runTurn(params);
+    // The REVIEWER chat call carries the digest in its user message.
+    const reviewerCall = chatCalls.find((c) =>
+      c.some((m) => (m.content ?? "").includes("Review the work-in-progress below")),
+    );
+    expect(reviewerCall).toBeTruthy();
+    expect(reviewerCall!.some((m) => (m.content ?? "").includes("harden the parser"))).toBe(true);
+  });
+});
+
+/* ── Compaction: efficient + accurate ─────────────────────────────── */
+
+describe("takeRecentTail", () => {
+  it("keeps the most recent messages and evicts an oversized older blob", () => {
+    const messages = [
+      toolResultMessage("t0", "y".repeat(5000)), // huge old tool output
+      userMessage("old request"),
+      assistantMessage("recent reply A"),
+      userMessage("recent reply B"),
+    ];
+    const { tail, evicted } = takeRecentTail(messages, 25, 2);
+    expect(evicted.length).toBe(2); // the blob + the old request
+    expect(tail.map((m) => m.content)).toEqual(["recent reply A", "recent reply B"]);
+  });
+
+  it("gives the whole history as tail when it is small enough", () => {
+    const messages = [userMessage("a"), assistantMessage("b")];
+    const { tail, evicted } = takeRecentTail(messages, 1000, 6);
+    expect(evicted).toEqual([]);
+    expect(tail.length).toBe(2);
+  });
+
+  it("does not let a huge tool output blow the tail past the min count", () => {
+    // minTail 2: newest two kept regardless; the huge tool body stays evicted.
+    const messages = [
+      toolResultMessage("t0", "x".repeat(50000)),
+      userMessage("keep1"),
+      assistantMessage("keep2"),
+    ];
+    const { tail, evicted } = takeRecentTail(messages, 8000, 2);
+    expect(tail.map((m) => m.content)).toEqual(["keep1", "keep2"]);
+    expect(evicted.length).toBe(1);
+  });
+});
+
+describe("reduceMessagesToDigest", () => {
+  it("stubs tool bodies entirely and previews text", () => {
+    const digest = reduceMessagesToDigest([
+      toolResultMessage("t0", "y".repeat(5000)),
+      userMessage("short request"),
+    ]);
+    expect(digest).toContain("- tool");
+    expect(digest).not.toContain("yyyy");
+    expect(digest).toContain("short request");
+  });
+
+  it("keeps user previews well above the shorter assistant cap", () => {
+    const long = "w".repeat(2000);
+    // User budget is 1500 chars — larger than the 400-char assistant cap, so
+    // a long user request survives far more of its body than a long assistant
+    // reply would (requests matter most for summary accuracy).
+    const digestUser = reduceMessagesToDigest([userMessage(long)]);
+    const digestAssistant = reduceMessagesToDigest([assistantMessage(long)]);
+    expect(digestUser.length).toBeGreaterThan(800);
+    expect(digestAssistant.length).toBeLessThan(500);
+  });
+
+  it("respects the hard input cap", () => {
+    const many = [];
+    for (let i = 0; i < 200; i++) many.push(userMessage("r".repeat(300)));
+    const digest = reduceMessagesToDigest(many);
+    expect(digest.length).toBeLessThanOrEqual(20_000 + 40);
+  });
+});
+
+describe("compactMessagesWithAi", () => {
+  it("keeps the recent tail and prepends an AI summary", async () => {
+    // Long enough to bypass the min-size guard, with an oversized OLD tool blob
+    // (should be evicted + stubbed, not kept verbatim).
+    const messages: ChatMessage[] = [
+      userMessage("old request one"),
+      assistantMessage("old response"),
+      toolResultMessage("t0", "big".repeat(4000)),
+    ];
+    for (let i = 0; i < 7; i++) messages.push(userMessage(`mid ${i}`));
+    messages.push(toolResultMessage("t1", "tail-result"));
+    messages.push(userMessage("newest request"));
+    messages.push(assistantMessage("newest response"));
+
+    const provider: ProviderService = {
+      async chat() {
+        return { message: { role: "assistant", content: "## Requests\n- old request one" }, usage: null };
+      },
+      async chatStream() {
+        return { message: { role: "assistant", content: null }, usage: null };
+      },
+    };
+    await compactMessagesWithAi(messages, provider);
+    // Summary first, recent tail preserved verbatim, tool blobs gone.
+    expect(messages[0]?.content).toContain("[AI Summary of Previous Conversation]");
+    expect(messages.some((m) => m.content === "newest request")).toBe(true);
+    expect(messages.some((m) => m.content === "newest response")).toBe(true);
+    expect(messages.some((m) => (m.content ?? "").includes("big".repeat(10)))).toBe(false);
+  });
+
+  it("degrades gracefully on summarizer failure, preserving the tail", async () => {
+    const messages = [
+      userMessage("old"),
+      toolResultMessage("t0", "x".repeat(100)),
+      userMessage("recent request"),
+    ];
+    const failing: ProviderService = {
+      async chat() {
+        throw new Error("network down");
+      },
+      async chatStream() {
+        return { message: { role: "assistant", content: null }, usage: null };
+      },
+    };
+    await compactMessagesWithAi(messages, failing);
+    // The recent message survives even when the LLM call fails.
+    expect(messages.some((m) => m.content === "recent request")).toBe(true);
+  });
+});

src/features/agent/application/turn_service.ts

@@ -8,6 +8,7 @@ import {
   type ToolCall,
   type ToolDef,
   type JsonValue,
+  Roles,
   systemMessage,
   userMessage,
   assistantMessage,
@@ -20,6 +21,7 @@ import {
   mainAgentPromptWithProjectContext,
   compactionPrompt,
   errorRecoveryNote,
+  reviewerPrompt,
 } from "@zesdex/agent";
 import type { ProviderService } from "./ports.ts";
 import type { ToolExecutor } from "./index.ts";
@@ -38,11 +40,92 @@ const PROJECT_CONTEXT_MAX_CHARS = 12_000;
 const RULE_FILENAMES = ["AGENTS.md", "agent.md", "CLAUDE.md", "claude.md", ".cursorrules", ".zesdexrules"];
 const COMPACT_KEEP_TAIL = 6;
 
+/** Recent-tail budget (chars): the most recent history kept verbatim on compact. */
+const COMPACT_TAIL_CHARS = 8_000;
+/** Caps what the summarizer actually receives, keeping input small and focused. */
+const COMPACT_MAX_INPUT_CHARS = 20_000;
+/** Preview budget for a user request inside the digest (requests matter most). */
+const COMPACT_USER_PREVIEW_CHARS = 1_500;
+/** Preview budget for assistant text inside the digest. */
+const COMPACT_TEXT_PREVIEW_CHARS = 400;
+
+/** Tools that mutate the filesystem — after these, a verify run is expected. */
+const MUTATOR_TOOLS = new Set(["write", "edit", "delete"]);
+
+/** Consecutive identical, non-progressing tool iterations before the loop stops. */
+const MAX_NO_PROGRESS_STREAK = 4;
+
+/** Self-review is bounded to this many passes per turn. */
+const MAX_REVIEW_PASSES = 1;
+
+/** Per-tool-call wall-clock budget before the tool is considered hung. */
+const TOOL_TIMEOUT_MS = 90_000;
+
 /** Whether the output string denotes a tool error. */
 function isErrorOutput(output: string): boolean {
   return output.startsWith("Error:");
 }
 
+/** Extract the numeric exit code from a `bash` tool result, or null if not a failure (0). */
+export function bashExitCode(output: string): number | null {
+  const match = output.match(/Exit code:\s*(\d+)/);
+  if (!match) return null;
+  const code = Number(match[1]);
+  return Number.isInteger(code) && code >= 0 ? code : null;
+}
+
+/**
+ * Whether a tool result is a real failure. `Error:` prefixes cover most tools;
+ * a `bash` command that exits non-zero returns an `Exit code: N` line instead.
+ */
+export function isToolFailure(toolName: string, output: string): boolean {
+  if (isErrorOutput(output)) return true;
+  if (toolName === "bash" || toolName === "bash_output") {
+    const code = bashExitCode(output);
+    if (code === null) return false; // no exit-code line → no signal
+    return code !== 0;
+  }
+  return false;
+}
+
+/**
+ * Best-effort derivation of the repository's verify command from convention
+ * manifests. Defaults to a safe lint+test invocation for Bun.
+ */
+export function deriveVerifyCommand(root: string): string {
+  const fs = require("node:fs");
+  // Bun project: prefer an explicit `check` (typecheck) script then test.
+  try {
+    const pkg = JSON.parse(fs.readFileSync(`${root}/package.json`, "utf8")) as {
+      scripts?: Record<string, string>;
+    };
+    const s = pkg?.scripts ?? {};
+    const parts: string[] = [];
+    if (s["check"] && typeof s["check"] === "string") parts.push(`bun run check`);
+    else if (s["typecheck"] && typeof s["typecheck"] === "string") parts.push(`bun run typecheck`);
+    if (s["lint"] && typeof s["lint"] === "string") parts.push(`bun run lint`);
+    if (s["test"] && typeof s["test"] === "string") parts.push(`bun run test`);
+    if (parts.length > 0) return parts.join(" && ");
+  } catch {
+    /* no package.json — fall through */
+  }
+  // Rust project.
+  try {
+    if (fs.existsSync(`${root}/Cargo.toml`)) return "cargo check && cargo test";
+  } catch {
+    /* ignore */
+  }
+  // Java/Maven.
+  try {
+    if (fs.existsSync(`${root}/pom.xml`) || fs.existsSync(`${root}/build.gradle`)) {
+      return "mvn test";
+    }
+  } catch {
+    /* ignore */
+  }
+  return "bun run check && bun test";
+}
+
 /** Truncate a long tool output, preserving the head + truncation marker. */
 export function truncateToolOutput(output: string): string {
   if (output.length <= TOOL_OUTPUT_MAX_CHARS) return output;
@@ -125,18 +208,32 @@ async function executeToolCall(
   executor: ToolExecutor,
   sink: TurnEventSink,
   tc: ToolCall,
+  signal?: AbortSignal,
 ): Promise<string> {
   const name = tc.function.name;
   const args = sanitizeToolArguments(tc.function.arguments);
 
-  let output: string;
+  const run = executor.execute(name, args as JsonValue).catch((e) => `Error: ${(e as Error).message}`);
-  try {
-    output = await executor.execute(name, args as JsonValue);
-  } catch (e) {
-    output = `Error: ${(e as Error).message}`;
-  }
 
-  const isError = isErrorOutput(output);
+  // Correct async cancellation: enforce a per-tool time budget and stop
+  // immediately if the turn is aborted while the tool is still running.
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  const output = await Promise.race([
+    run,
+    new Promise<string>((resolve) => {
+      const onAbort = () => resolve("Error: Turn aborted by user");
+      timer = setTimeout(() => resolve(`Error: Tool timed out after ${TOOL_TIMEOUT_MS / 1000}s`), TOOL_TIMEOUT_MS);
+      signal?.addEventListener("abort", onAbort, { once: true });
+      // Release the abort listener + timer once the tool settles either way.
+      void run.finally(() => {
+        clearTimeout(timer);
+        signal?.removeEventListener("abort", onAbort);
+      });
+    }),
+  ]);
+  if (timer) clearTimeout(timer);
+
+  const isError = isToolFailure(name, output);
   const truncated = truncateToolOutput(output);
 
   sink.push({
@@ -156,6 +253,7 @@ async function executeToolCallsInParallel(
   executor: ToolExecutor,
   sink: TurnEventSink,
   toolCalls: ToolCall[],
+  signal?: AbortSignal,
 ): Promise<string[]> {
   // Simple bounded concurrency preserving input order.
   const results: string[] = new Array(toolCalls.length);
@@ -165,7 +263,7 @@ async function executeToolCallsInParallel(
     while (true) {
       const idx = next++;
       if (idx >= toolCalls.length) return;
-      results[idx] = await executeToolCall(executor, sink, toolCalls[idx]!);
+      results[idx] = await executeToolCall(executor, sink, toolCalls[idx]!, signal);
     }
   }
 
@@ -179,8 +277,75 @@ async function executeToolCallsInParallel(
 /* -------------------------------------------------------------------------- */
 
 /**
- * Compact oversized conversation history using AI summarisation. At most once
+ * Split history into a recent tail (kept verbatim, bound by CHAR budget so
- * per turn. Keeps the last COMPACT_KEEP_TAIL messages.
+ * huge tool outputs don't monopolise it) and the evicted prefix to summarize.
+ */
+export function takeRecentTail(
+  messages: ChatMessage[],
+  tailChars = COMPACT_TAIL_CHARS,
+  minTail = COMPACT_KEEP_TAIL,
+): { tail: ChatMessage[]; evicted: ChatMessage[] } {
+  if (messages.length <= minTail) return { tail: [...messages], evicted: [] };
+  let used = 0;
+  let keep = 0;
+  // Walk from the newest message backward. Always keep at least minTail; then
+  // stop once the aggregated char budget is exceeded.
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const len = messages[i]!.content?.length ?? 0;
+    if (keep >= minTail && used + len > tailChars) break;
+    keep += 1;
+    used += len;
+  }
+  const tail = messages.slice(messages.length - keep);
+  const evicted = messages.slice(0, messages.length - keep);
+  return { tail, evicted };
+}
+
+/**
+ * Reduce older history to a compact digest for the summarizer. Tool-result
+ * bodies are dropped entirely (they are noise for a summary); user requests
+ * get a generous preview; assistant text gets a shorter one. Caps the result
+ * at COMPACT_MAX_INPUT_CHARS so the summarizer sees a small, focused input.
+ */
+export function reduceMessagesToDigest(messages: ChatMessage[]): string {
+  let out = "";
+  let budget = COMPACT_MAX_INPUT_CHARS;
+  for (const m of messages) {
+    if (budget <= 0) break;
+    if (m.role === Roles.Tool) {
+      const name = m.name ?? "tool";
+      const line = `- tool ${name} executed\n`;
+      out += line;
+      budget -= line.length;
+      continue;
+    }
+    const text = m.content?.trim() ?? "";
+    if (text === "") continue; // assistant messages that only carried tool calls
+    const cap = m.role === Roles.User ? COMPACT_USER_PREVIEW_CHARS : COMPACT_TEXT_PREVIEW_CHARS;
+    let preview = text.replace(/\s*\n+\s*/g, " ");
+    if (preview.length > cap) {
+      preview = `${preview.slice(0, cap)}…[+${text.length - cap} ch]`;
+    }
+    const line = `- ${m.role}: ${preview}\n`;
+    out += line;
+    budget -= line.length;
+  }
+  // Hard guarantee: never exceed the summarizer input cap.
+  if (out.length > COMPACT_MAX_INPUT_CHARS) return out.slice(0, COMPACT_MAX_INPUT_CHARS);
+  return out.trim();
+}
+
+/**
+ * Compact oversized conversation history efficiently yet accurately.
+ *
+ *  - Only the OLDER history is summarized; the recent tail is kept verbatim
+ *    (bounded by char budget, not message count, so tool bloat can't win the
+ *    budget).
+ *  - The evicted history is REDUCED to a small digest before the summarizer
+ *    sees it, so the LLM works on a focused input (cheaper + more accurate)
+ *    instead of a ~60k-char dump.
+ *  - On failure it degrades gracefully: keeps the recent tail and a marker,
+ *    rather than wiping context to a useless sentinel.
  */
 export async function compactMessagesWithAi(
   messages: ChatMessage[],
@@ -188,17 +353,35 @@ export async function compactMessagesWithAi(
 ): Promise<void> {
   if (messages.length <= COMPACT_KEEP_TAIL + 2) return;
 
-  const splitIdx = messages.length - COMPACT_KEEP_TAIL;
+  const { tail, evicted } = takeRecentTail(messages);
-  const evicted = messages.splice(0, splitIdx);
+  const digest = reduceMessagesToDigest(evicted);
+  if (digest.trim() === "") {
+    // Nothing worth summarizing (e.g. only stubbed tool bodies) — keep it all.
+    return;
+  }
 
-  const summaryPrompt: ChatMessage[] = [systemMessage(compactionPrompt()), ...evicted, userMessage("Please summarise our previous conversation above for context continuity.")];
+  const summaryPrompt: ChatMessage[] = [
+    systemMessage(compactionPrompt()),
+    userMessage(
+      "The conversation below is the OLDER part of a session. Recent messages are kept separately, so do not preserve them. Compress the older part into the requested summary format.\n\n---\n" + digest,
+    ),
+  ];
 
   try {
     const { message } = await provider.chat(summaryPrompt, undefined, 1024, 0.3);
-    const summaryText = message.content ?? "Previous context summarised.";
+    const summaryText = message.content?.trim() ?? "";
-    messages.unshift(systemMessage(`[AI Summary of Previous Conversation]\n${summaryText.trim()}`));
+    const rebuilt: ChatMessage[] = [];
+    if (summaryText !== "") {
+      rebuilt.push(systemMessage(`[AI Summary of Previous Conversation]\n${summaryText}`));
+    } else {
+      rebuilt.push(systemMessage("[Earlier conversation summarized]"));
+    }
+    rebuilt.push(...tail);
+    messages.splice(0, messages.length, ...rebuilt);
   } catch {
-    messages.unshift(systemMessage("[Earlier conversation messages compacted to save context window]"));
+    // Graceful degradation: never wipe recent state on a summarizer failure.
+    const rebuilt: ChatMessage[] = [systemMessage("[Earlier context compressed; recent messages follow.]"), ...tail];
+    messages.splice(0, messages.length, ...rebuilt);
   }
 }
 
@@ -268,9 +451,11 @@ export class AgentTurnServiceImpl {
     const abort = params.abort;
     const { in_flight } = params;
 
-    // Insert system prompt at index 0 with repo conventions loaded.
+    // Insert system prompt at index 0 with repo conventions + verify command.
-    const projectContext = buildProjectContext(params.workspace_roots[0] ?? ".");
+    const root = params.workspace_roots[0] ?? ".";
-    const systemPrompt = mainAgentPromptWithProjectContext(projectContext);
+    const projectContext = buildProjectContext(root);
+    const verifyCommand = deriveVerifyCommand(root);
+    const systemPrompt = mainAgentPromptWithProjectContext(projectContext, verifyCommand);
     params.messages.unshift(systemMessage(systemPrompt));
     const originalCount = params.messages.length;
 
@@ -280,6 +465,12 @@ export class AgentTurnServiceImpl {
 
     const errors = new ErrorTracker();
     let sawToolCalls = false;
+    let sawMutator = false;
+    let pendingVerify = false;
+    let verifyPrompted = false;
+    let noProgressStreak = 0;
+    let lastSignature: string | null = null;
+    let reviewPasses = 0;
 
     for (let iteration = 0; iteration < MAX_TURN_ITERATIONS; iteration++) {
       // Check abort flag.
@@ -293,12 +484,33 @@ export class AgentTurnServiceImpl {
         break;
       }
 
+      // Convergence guard: bail out of a loop stuck re-issuing the same call.
+      if (noProgressStreak >= MAX_NO_PROGRESS_STREAK) {
+        this.push(sink, {
+          kind: "system_note",
+          systemKind: "warn",
+          message: "Stopping: the same tool call is being repeated without any progress.",
+        });
+        break;
+      }
+
+      // Verify-after-edit: nudge the model to run the check before concluding.
+      if (pendingVerify && !verifyPrompted) {
+        params.messages.push(
+          systemMessage(
+            `[Flow] You just modified files. Run the verify command via \`bash\` now (${verifyCommand}) and resolve any failures before concluding your turn.`,
+          ),
+        );
+        verifyPrompted = true;
+      }
+
       // Auto-compact oversized history before the LLM call.
       await this.autoCompactIfNeeded(params.messages);
 
-      // Adaptive generation parameters.
+      // Generation parameters: a config override wins, else fall back to the
-      const maxTokens = adaptiveMaxTokens(requestLen);
+// adaptive defaults (lower temperature once the model is doing tool work).
-      const temperature = sawToolCalls ? 0.2 : 0.7;
+      const maxTokens = params.max_tokens ?? adaptiveMaxTokens(requestLen);
+      const temperature = params.temperature ?? (sawToolCalls ? 0.2 : 0.7);
 
       this.push(sink, { kind: "stream_start" });
 
@@ -336,19 +548,52 @@ export class AgentTurnServiceImpl {
       const parallel = toolCalls.length > 1 && toolCalls.every(isParallelSafe);
 
       const outputs = parallel
-        ? await executeToolCallsInParallel(this.toolExecutor, sink, toolCalls)
+        ? await executeToolCallsInParallel(this.toolExecutor, sink, toolCalls, abort.signal)
         : await (async () => {
             const seq: string[] = [];
-            for (const tc of toolCalls) seq.push(await executeToolCall(this.toolExecutor, sink, tc));
+            for (const tc of toolCalls) seq.push(await executeToolCall(this.toolExecutor, sink, tc, abort.signal));
             return seq;
           })();
 
+      let mutated = false;
+      let verified = false;
+      let batchSignature = "";
       for (let i = 0; i < toolCalls.length; i++) {
         const tc = toolCalls[i]!;
         const output = outputs[i]!;
-        if (isErrorOutput(output)) errors.record(tc.function.name, output, params.messages);
+        if (isToolFailure(tc.function.name, output)) errors.record(tc.function.name, output, params.messages);
+        if (MUTATOR_TOOLS.has(tc.function.name)) {
+          mutated = true;
+          sawMutator = true;
+        }
+        if ((tc.function.name === "bash" || tc.function.name === "bash_output") && bashExitCode(output) === 0) {
+          verified = true;
+        }
+        // Batch signature: concat of tool+output for no-progress detection.
+        batchSignature += `${tc.function.name}${output.length}`;
         params.messages.push(toolResultMessage(tc.id, output));
       }
+
+      // No-progress: byte-identical batch (same tools, same output lengths) → streak.
+      if (toolCalls.length === 1 && batchSignature === lastSignature) noProgressStreak += 1;
+      else noProgressStreak = 0;
+      lastSignature = batchSignature;
+
+      if (mutated) {
+        pendingVerify = true;
+        // Do NOT re-arm verifyPrompted: the verify instruction is already in
+        // context, so re-injecting it on every edit batch only bloats tokens.
+      }
+      if (verified) {
+        pendingVerify = false; // a successful bash run satisfies the verify nudge
+        verifyPrompted = false; // a later edit batch may nudge again
+      }
+
+      // Bounded self-review pass after file mutations (at most one per turn).
+      if (sawMutator && reviewPasses < MAX_REVIEW_PASSES && !abort.signal.aborted) {
+        reviewPasses += 1;
+        await this.reviewWork(params.messages, abort, sink);
+      }
     }
 
     // Remove the synthetic sys_msg before emitting to the transcript.
@@ -357,4 +602,37 @@ export class AgentTurnServiceImpl {
     this.push(sink, { kind: "done" });
     in_flight.value = false;
   }
+
+  /** Run one bounded self-review of the work so far and feed critique back. */
+  private async reviewWork(
+    messages: ChatMessage[],
+    abort: AbortController,
+    sink: TurnEventSink,
+  ): Promise<void> {
+    if (abort.signal.aborted) return;
+    // Reduce the work-in-progress to a focused digest so the reviewer has real
+    // context to critique (not an empty slate that invites hallucinated issues).
+    const digest = reduceMessagesToDigest(messages.slice(-80));
+    if (digest.trim() === "") return;
+    try {
+      const result = await this.provider.chat(
+        [
+          systemMessage(reviewerPrompt()),
+          userMessage(`Review the work-in-progress below and list actionable issues:\n\n---\n${digest}`),
+        ],
+        undefined,
+        1024,
+        0.3,
+      );
+      const critique = result.message.content ?? "";
+      if (result.usage) {
+        this.push(sink, { kind: "review_usage", tokens_in: result.usage[0], tokens_out: result.usage[1] });
+      }
+      if (critique.trim() === "" || critique.trim().toUpperCase() === "NO ISSUES FOUND") return;
+      // Feed the critique back as a system message so the next iteration resolves it.
+      messages.push(systemMessage(`[Reviewer]\n${critique.trim()}`));
+    } catch {
+      /* review is best-effort — non-fatal */
+    }
+  }
 }

src/features/agent/domain/mod.ts

@@ -62,7 +62,6 @@ export function agentStatusDisplay(s: AgentStatus, error?: string): string {
 
 /** Events emitted onto the turn-event queue while an agent turn runs. */
 export type TurnEvent =
-  | { kind: "assistant_message"; message: ChatMessage }
   | { kind: "tool_result"; tool_call_id: string; tool_name: string; output: string; is_error: boolean; path: string | null }
   | { kind: "system_note"; systemKind: string; message: string }
   | { kind: "stream_start" }
@@ -183,6 +182,10 @@ export interface AgentTurnParams {
   api_key: string;
   model: string;
   api_base?: string;
+  /** Optional override for the LLM max_tokens ceiling (defaults to adaptive). */
+  max_tokens?: number;
+  /** Optional override for the LLM temperature (defaults to 0.7 / 0.2 tooling). */
+  temperature?: number;
 }
 
 /** Minimal event-sink abstraction (backs the Rust `Arc<Mutex<VecDeque>>`). */

src/features/agent/domain/prompt.test.ts

@@ -0,0 +1,105 @@
+/**
+ * Tests pinning the rewritten prompts to be consistent with the REAL tool
+ * inventory: they must never reference the phantom `explore_codebase`, must
+ * list real tool names, and must carry the concrete verify command.
+ */
+import { describe, expect, test } from "bun:test";
+import {
+  mainAgentPrompt,
+  mainAgentPromptWithProjectContext,
+  subagentDirective,
+  compactionPrompt,
+  reviewerPrompt,
+  consensusSynthesizerPrompt,
+  errorRecoveryNote,
+} from "./prompt.ts";
+
+/** All real tool names (mirrors registry.allTools()). */
+const REAL_TOOLS = new Set([
+  "read", "write", "edit", "delete",
+  "grep", "glob", "semantic_search", "rebuild_index", "list_symbols",
+  "bash", "bash_output", "bash_kill", "git_operator", "git_worktree", "git_cred",
+  "cd", "dir_list", "dir_cache_update",
+  "plan_enter", "plan_ready", "sequential_think", "todowrite", "todofinish",
+  "workflow_run", "hive_mind", "note_finding", "read_findings",
+  "spawn_agents", "spawn_pipeline", "parallel_delegate",
+  "remember", "forget", "recall",
+  "web_search", "best_practice", "commit_convention", "pong",
+]);
+
+describe("prompt tool-consistency", () => {
+  test("main prompt never references non-existent explore_codebase", () => {
+    expect(mainAgentPrompt()).not.toContain("explore_codebase");
+  });
+
+  test("every backticked tool token in the main prompt is a real tool", () => {
+    const s = mainAgentPrompt();
+    const backticks = s.match(/`([a-z_]+)`/g) ?? [];
+    const tokens = backticks.map((t) => t.slice(1, -1));
+    // `plan` and `verify` appear in backticks as prose (a command/word), not tools.
+    const proseWords = new Set(["plan", "verify", "check"]);
+    const ghosts = tokens.filter((t) => !REAL_TOOLS.has(t) && !proseWords.has(t));
+    expect(ghosts).toEqual([]);
+  });
+
+  test("main prompt anchors the key real tools", () => {
+    const s = mainAgentPrompt();
+    for (const tool of ["plan_enter", "todowrite", "workflow_run", "spawn_agents", "hive_mind", "recall", "remember"]) {
+      expect(s).toContain(tool);
+    }
+  });
+
+  test("verify command is injected as a concrete line when provided", () => {
+    const s = mainAgentPrompt("bun run check && bun test");
+    expect(s).toContain("VERIFY COMMAND");
+    expect(s).toContain("bun run check && bun test");
+  });
+
+  test("without a verify command, no VERIFY COMMAND block is emitted", () => {
+    expect(mainAgentPrompt()).not.toContain("VERIFY COMMAND");
+  });
+
+  test("project-context variant appends the context block", () => {
+    const s = mainAgentPromptWithProjectContext("## Rules\nno threads", "bun test");
+    expect(s).toContain("## PROJECT CONTEXT");
+    expect(s).toContain("no threads");
+    expect(s).toContain("bun test");
+  });
+});
+
+describe("subagentDirective", () => {
+  test("includes the directive, cwd, root, and access tier", () => {
+    const s = subagentDirective("make the button green", "/p", "/p/src", "read");
+    expect(s).toContain("make the button green");
+    expect(s).toContain("/p");
+    expect(s).toContain("READ-ONLY");
+  });
+
+  test("no-tier call defaults to full access", () => {
+    const s = subagentDirective("do it", "/p", "/p");
+    expect(s).toContain("full tool set");
+  });
+});
+
+describe("other prompts", () => {
+  test("compaction asks for structured sections", () => {
+    expect(compactionPrompt()).toContain("## Requests");
+    expect(compactionPrompt()).toContain("## Work done");
+  });
+
+  test("reviewer emits NO ISSUES FOUND sentinel contract", () => {
+    expect(reviewerPrompt()).toContain("NO ISSUES FOUND");
+  });
+
+  test("consensus synthesizer covers the four sections", () => {
+    const s = consensusSynthesizerPrompt();
+    expect(s).toContain("AGREEMENTS");
+    expect(s).toContain("CONFLICTS");
+    expect(s).toContain("KEY FINDINGS");
+    expect(s).toContain("RECOMMENDATION");
+  });
+
+  test("error recovery note is present", () => {
+    expect(errorRecoveryNote("read", "nope")).toContain("[System note]");
+  });
+});

src/features/agent/domain/prompt.ts

@@ -1,32 +1,54 @@
-/** System prompts and directive templates. Mirrors `agent/prompt.rs`. */
+/**
+ * System prompts and directive templates. Mirrors `agent/prompt.rs`.
+ *
+ * Every string here is written against the REAL built-in tool inventory
+ * (see `infrastructure/tools/registry.ts`). There is intentionally no
+ * reference to a hypothetical `explore_codebase`; orientation is done through
+ * the actual `read` / `grep` / `glob` / `semantic_search` tools.
+ */
+
+const MUTATOR_TOOLS = "`write`, `edit`, `delete`";
+
+/** Shared operating rules referenced by main and sub-agent prompts. */
+const SHARED_DISCIPLINE = `TOOL DISCIPLINE:
+- Tool failures arrive prefixed with \`Error:\`. A \`bash\` command that exits non-zero returns an \`Exit code: N\` line (NOT an \`Error:\` prefix) — treat a non-zero exit the same as a failure.
+- Read before you write: never edit a file you have not read first.
+- Make the smallest correct change. Do not rewrite large sections unless the task explicitly calls for it.
+- Do not re-read files whose contents are already in your context. Prefer targeted \`grep\`/\`read\` over broad scans (\`glob\`, \`semantic_search\`) once you know where the code lives.
+- Follow the repository conventions from the PROJECT CONTEXT block when present.`;
 
 /** Build the main-agent system prompt. */
-export function mainAgentPrompt(): string {
+export function mainAgentPrompt(verifyCommand?: string): string {
-  return `You are Zesdex, an AI coding assistant. You have access to various tools via native function calling to help the user.
+  return `You are Zesdex, an autonomous AI coding agent. You help by doing real work against the user's repository using native function-calling tools.
 
-TOKEN BUDGET — BE EFFICIENT:
+HOW TO WORK (choose the lightest sufficient path):
-- For simple/factual questions, answer directly. Do NOT call tools.
+1. Simple or factual: answer directly. Do NOT call tools.
-- For complex or unfamiliar code tasks, call \`explore_codebase\` ONCE at the start to locate relevant code, then work from that context.
+2. Needs code context: orient ONCE with \`read\`/\`grep\`/\`glob\` (or \`semantic_search\` for fuzzy questions), then work from that context.
-- Keep tool usage minimal: prefer \`grep\`/\`glob\`/\`read\` for targeted lookups; avoid re-reading files you already have in context.
+3. Multi-step or complex: enter a plan with \`plan_enter\`, track steps with \`todowrite\`, then execute.
-- Keep responses concise; do not repeat tool output verbatim.
+4. Genuinely parallel or multi-angle exploration: prefer \`workflow_run\`, \`spawn_agents\`, \`parallel_delegate\`, or \`hive_mind\`. Do NOT spawn agents for work you can do directly in a few steps.
 
-CRITICAL DIRECTIVES & PRIORITY HIERARCHY:
+${SHARED_DISCIPLINE}
-1. WORKFLOW FIRST: For any multi-step, complex, or non-trivial task, you MUST prioritise using \`workflow_run\` (to construct and execute a multi-phase YAML workflow) or \`hive_mind\` (to orchestrate parallel autonomous agents). Workflows are your primary strategy.
-2. PLANNING & TODOs: Use \`plan_enter\` to establish high-level architectural plans and \`todowrite\` to maintain granular task checklists.
-3. REASONING: Use \`seq_think\` for deep step-by-step analysis.
-4. TOOL EXECUTION: Execute individual tools (file edits, terminal commands) within or guided by your workflows. If an error occurs, analyse and fix it.
 
-VERIFY AFTER EDIT (CLAUDE-CODE STYLE):
+VERIFICATION CONTRACT:
-- After modifying code (edit/write), run the repo's check command via \`bash\` before ending the turn: \`cargo check\` / \`cargo clippy\` / \`cargo test\` for Rust, or the equivalent lint/test (\`bun run lint && bun run test\`, \`npm test\`, etc.) for other stacks. Pick the project's actual verify command (see PROJECT CONTEXT / AGENTS.md when present).
+- After any of ${MUTATOR_TOOLS}, run the repository's verify command below and resolve failures before concluding your turn.
-- If the check fails, fix the errors you can see and re-run; only end the turn after the check passes or you cannot resolve a failure yourself (then report it explicitly).
+- Only claim code compiles or tests pass after running a real check.
-- Do NOT claim code compiles or works without running a real check.
+${verifyCommand ? `\nVERIFY COMMAND (run via \`bash\`):\n${verifyCommand}` : ""}
 
-Respond conversationally, concisely, and helpfully.`;
+OUTPUT CONTRACT:
+- End with a concise summary of what you did and an explicit next step. Do not dump raw tool output into your reply.
+- If you cannot complete something, state exactly what is blocking you.
+- Reply in the same language the user writes in.
+
+MEMORY:
+- At the start of a task, \`recall\` memory that may be relevant; use \`remember\` to store any hard-won insight or lesson worth keeping.`;
 }
 
-/** Main-agent prompt with an injected `## PROJECT CONTEXT` block. Empty context → base prompt. */
+/** Main-agent prompt with an optional project-context block and verify command. */
-export function mainAgentPromptWithProjectContext(projectContext: string): string {
+export function mainAgentPromptWithProjectContext(
-  const base = mainAgentPrompt();
+  projectContext: string,
+  verifyCommand?: string,
+): string {
+  const base = mainAgentPrompt(verifyCommand);
   const context = projectContext.trim();
   if (context === "") return base;
   return `${base}
@@ -35,29 +57,51 @@ export function mainAgentPromptWithProjectContext(projectContext: string): strin
 ${context}`;
 }
 
-/** Build a subagent directive prompt. */
+/** Build a subagent directive prompt. Access tier tells the subagent its limits. */
-export function subagentDirective(directive: string, cwd: string, wsRoot: string): string {
+export function subagentDirective(
-  return `You are a focused subagent.
+  directive: string,
+  cwd: string,
+  wsRoot: string,
+  access: "read" | "write" | "full" = "full",
+): string {
+  const tierNote =
+    access === "read"
+      ? "You have READ-ONLY tools: you may inspect files and search, but must NOT modify anything."
+      : access === "write"
+        ? "You have WRITE tools (file edits) but NOT shell or network execution."
+        : "You have the full tool set.";
+  return `You are a focused subagent working autonomously on a single directive.
 
 Current directory (PWD): ${cwd}
 Workspace root: ${wsRoot}
+Access tier: ${access} — ${tierNote}
 
-Your directive:
+${SHARED_DISCIPLINE}
+
+Your directive (source of truth — follow it exactly):
 ${directive}
 
-Complete the directive autonomously using the tools available to you. Return your final answer when done.`;
+Complete the directive autonomously with the tools available to you. When done, return a concise final answer that reports what you changed (or found) and any open issues. Do not ask for permission; act within your access tier.`;
 }
 
 /** Build a conversation-compaction prompt. */
 export function compactionPrompt(): string {
-  return `You are a helpful assistant summarising conversation history. Provide a concise summary of the key user requests, decisions, tools executed, and modified files. Format as a clear bulleted list.`;
+  return `You are condensing a long conversation to preserve context continuity without losing important detail.
+
+Produce a concise markdown recap with these sections (skip any section that has no content):
+- ## Requests: what the user asked for.
+- ## Decisions: important choices made and why.
+- ## Work done: tools executed and files created or modified (with paths).
+- ## Open threads: unresolved items, pending questions, or what is still to do.
+
+Be precise about file paths and current state. Do not invent details not present in the conversation.`;
 }
 
 /** Directive for a lightweight context-scout subagent. */
 export function exploreScoutDirective(): string {
   return `You are a codebase context scout. Given the workspace root, quickly locate the code that is most relevant to the user's request:
-1. Run semantic_search once with the user's key terms.
+1. Search for the relevant symbols/terms with \`grep\` or \`glob\`; use \`semantic_search\` only when the query is fuzzy.
-2. Read up to the 3 most relevant files (use grep for symbols if needed).
+2. Read up to the 3 most relevant files.
 3. Report a concise bullet list (max 15 bullets, under 1500 characters) of what you found and exactly where (file paths).
 Do NOT rebuild the index. Do NOT enumerate unrelated files. Be brief.`;
 }
@@ -65,4 +109,31 @@ Do NOT rebuild the index. Do NOT enumerate unrelated files. Be brief.`;
 /** System note injected after repeated tool errors. */
 export function errorRecoveryNote(toolName: string, lastError: string): string {
   return `[System note] The tool \`${toolName}\` failed repeatedly with: "${lastError}". Try an alternative approach (verify paths, correct arguments, use a different tool, or finish without this tool). Do NOT retry the same call.`;
-}
+}
+
+/** System prompt for the optional single-pass self-review after edits. */
+export function reviewerPrompt(): string {
+  return `You are a code reviewer for an autonomous agent's work-in-progress. You are shown the conversation and the changes made so far.
+
+Critique the work for:
+- CORRECTNESS: would the changes work as intended? any obvious bugs, edge cases, or missing pieces?
+- SPEC COMPLIANCE: did the agent actually do what was asked, or did it drift?
+- QUALITY: are the changes minimal, conventional, and consistent with the repository's style?
+
+Return a tight list of the most important ACTIONABLE issues only, each as:
+- [PRIORITY: high|medium] <one concrete fix>, e.g. "high: handle empty input in parse()" or "medium: run the tests; they reference a removed export".
+
+Do NOT praise the work. Do NOT restate the plan. If the work is already correct and complete, return exactly the single line "NO ISSUES FOUND". Be concrete enough that a next iteration can act without re-deriving context.`;
+}
+
+/** System prompt for reconciling multi-node hive-mind outputs into one consensus. */
+export function consensusSynthesizerPrompt(): string {
+  return `You are a consensus synthesizer for a multi-agent hive mind. Several independent nodes analysed a problem and produced the outputs below. Distill them into ONE coherent consensus report with these sections:
+- AGREEMENTS: points multiple nodes converge on.
+- CONFLICTS: contradictory conclusions, with which node(s) support each side.
+- KEY FINDINGS: the most important, actionable takeaways.
+- RECOMMENDATION: a single recommended next action, or 'no clear consensus' if the outputs are too divergent.
+
+Be concise and factual. If a node errored, note it and ignore its content.
+Do not invent facts not present in the node outputs.`;
+}

src/features/agent/infrastructure/llm/provider.test.ts

@@ -0,0 +1,52 @@
+/**
+ * Tests for LlmClient.listModels() — the live /models fetch backing the TUI
+ * model picker. Mocks global fetch so no network is touched.
+ */
+import { describe, test, expect, mock } from "bun:test";
+import { LlmClient } from "./provider.ts";
+
+/** Replace global fetch once, returning the given JSON body for /models. */
+async function listWith(fetchImpl: (url: string) => Promise<Response>): Promise<string[]> {
+  const orig = globalThis.fetch;
+  globalThis.fetch = fetchImpl as typeof fetch;
+  try {
+    const client = new LlmClient("key", "some-model", "https://api.example.com/v1");
+    return await client.listModels();
+  } finally {
+    globalThis.fetch = orig;
+  }
+}
+
+describe("LlmClient.listModels", () => {
+  test("parses OpenAI-style { data: [{id}] }", async () => {
+    const rows = await listWith(async (url: string) => {
+      expect(url).toBe("https://api.example.com/v1/models");
+      return new Response(JSON.stringify({ data: [{ id: "gpt-5" }, { id: "deepseek-v3" }] }), {
+        status: 200,
+        headers: { "Content-Type": "application/json" },
+      });
+    });
+    expect(rows).toEqual(["gpt-5", "deepseek-v3"]);
+  });
+
+  test("parses { models: [string] } variant", async () => {
+    const rows = await listWith(async () =>
+      new Response(JSON.stringify({ models: ["a", "b"] }), { status: 200 }),
+    );
+    expect(rows).toEqual(["a", "b"]);
+  });
+
+  test("filters out missing ids", async () => {
+    const rows = await listWith(async () =>
+      new Response(JSON.stringify({ data: [{ id: "x" }, {}, { id: "y" }] }), { status: 200 }),
+    );
+    expect(rows).toEqual(["x", "y"]);
+  });
+
+  test("returns [] on non-OK or throw", async () => {
+    const rows = await listWith(async () => new Response("forbidden", { status: 403 }));
+    expect(rows).toEqual([]);
+  });
+});
+
+void mock;

src/features/agent/infrastructure/llm/provider.ts

@@ -123,6 +123,35 @@ export class LlmClient implements ProviderService {
     return `API request failed: ${err?.message ?? String(e)}`;
   }
 
+  /**
+   * List the model ids exposed by the provider's `/models` endpoint
+   * (OpenAI/Anthropic-compatible). Returns an empty array when the endpoint
+   * isn't exposed or the call fails — callers fall back to static catalogs.
+   */
+  async listModels(): Promise<string[]> {
+    const url = `${this.baseUrl}/models`;
+    try {
+      const resp = await fetch(url, {
+        method: "GET",
+        headers: this.headers(),
+        signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),
+      });
+      if (!resp.ok) return [];
+      const data = (await resp.json()) as {
+        data?: { id?: string }[] | null;
+        models?: (string | { id?: string })[] | null;
+      };
+      const fromData = (data.data ?? []).map((m) => m?.id ?? "").filter(Boolean);
+      if (fromData.length) return fromData;
+      const fromModels = (data.models ?? [])
+        .map((m) => (typeof m === "string" ? m : m?.id ?? ""))
+        .filter(Boolean);
+      return fromModels;
+    } catch {
+      return [];
+    }
+  }
+
   async chat(
     messages: ChatMessage[],
     tools?: ToolDef[],

src/features/agent/infrastructure/tools/parallel_delegate.test.ts

@@ -0,0 +1,35 @@
+/**
+ * Tests for the LLM-driven parallel-delegation splitter. The decomposition is
+ * decided by the AI (no hardcoded keyword guessing); these tests lock in the
+ * tolerant JSON parsing and the access-tier normalization.
+ */
+import { describe, expect, test } from "bun:test";
+import { parseDirectivesJson } from "./parallel_delegate.ts";
+
+describe("parseDirectivesJson", () => {
+  test("parses a clean JSON array", () => {
+    const out = parseDirectivesJson(
+      `[{"directive":"Scan the config code","access":"read"},{"directive":"Implement the change","access":"full"}]`,
+    );
+    expect(out.length).toBe(2);
+    expect(out[0]).toEqual({ directive: "Scan the config code", access: "read" });
+  });
+
+  test("extracts the array from markdown fences with a preamble", () => {
+    const out = parseDirectivesJson(
+      'Here are the subtasks:\n```json\n[{"directive":"A","access":"read"},{"directive":"B","access":"full"}]\n```\nDone.',
+    );
+    expect(out.map((d) => d.directive)).toEqual(["A", "B"]);
+  });
+
+  test("normalises unknown access tiers to write", () => {
+    const out = parseDirectivesJson(`[{"directive":"do it","access":"admin"}]`);
+    expect(out[0]!.access).toBe("write");
+  });
+
+  test("drops empty directives and malformed input", () => {
+    expect(parseDirectivesJson(`[{"directive":"","access":"write"}]`)).toEqual([]);
+    expect(parseDirectivesJson("not json at all")).toEqual([]);
+    expect(parseDirectivesJson("")).toEqual([]);
+  });
+});

src/features/agent/infrastructure/tools/parallel_delegate.ts

@@ -50,7 +50,8 @@ export class ParallelDelegate implements Tool {
         }))
         .filter((d) => d.directive !== "");
     } else {
-      directives = fallbackSplit(task, maxParallel);
+      // Let the LLM decide the decomposition instead of guessing keywords.
+      directives = await aiDecomposeTask(task, maxParallel);
     }
 
     if (directives.length === 0) {
@@ -73,22 +74,67 @@ export class ParallelDelegate implements Tool {
   }
 }
 
-/** Fallback splitting when LLM is unavailable. */
+/**
-function fallbackSplit(task: string, maxParallel: number): Array<{ directive: string; access: string }> {
+ * Have the LLM decompose a task into focused, non-overlapping parallel
-  const directives: Array<{ directive: string; access: string }> = [];
+ * sub-directives. This is the AI deciding how to split the work — no keyword
-  if (task.includes("backend") || task.includes("api") || task.includes("server")) {
+ * guessing. On LLM failure (no provider / error), degrades to a single honest
-    directives.push({ directive: `Implement the backend/API components for: ${task}`, access: "write" });
+ * full-task directive rather than fabricated "part N" stubs.
+ */
+async function aiDecomposeTask(
+  task: string,
+  maxParallel: number,
+): Promise<Array<{ directive: string; access: string }>> {
+  try {
+    const { resolveConfig } = await import("../../../subagent/infrastructure/config_resolver.ts");
+    const { buildProviderService } = await import("../../../subagent/infrastructure/http_provider.ts");
+    const { baseUrl, apiKey, model } = await resolveConfig();
+    const svc = buildProviderService(baseUrl, apiKey, model);
+
+    const systemPrompt =
+      "You decompose a large task into independent, non-overlapping subtasks that can be worked on " +
+      "in parallel. Return a JSON array of up to the requested count of objects, each with a " +
+      "'directive' (a concrete, self-contained action) and an 'access' tier of exactly one of " +
+      "read | write | full. read = inspect/search only; write = edit files; full = edit + run shell. " +
+      "Prefer fewer, genuinely parallel directives over many that touch the same files. " +
+      "Return ONLY the JSON array, no prose.";
+
+    const { message } = await svc.chat(
+      [
+        { role: "system", content: systemPrompt },
+        { role: "user", content: `Task: ${task}\nMax subtasks: ${maxParallel}\n` },
+      ],
+      undefined,
+      1024,
+      0.3,
+    );
+
+    const parsed = parseDirectivesJson(message.content ?? "");
+    if (parsed.length > 0) return parsed.slice(0, maxParallel);
+  } catch {
+    /* fall through to the single-directive fallback */
+  }
+  // Honest fallback: run the whole task as one agent rather than guessing a split.
+  return [{ directive: task, access: "full" }];
+}
+
+/** Parse an LLM JSON array of directives, tolerant of stray prose/markdown fences. */
+export function parseDirectivesJson(raw: string): Array<{ directive: string; access: string }> {
+  const fenced = raw.match(/```(?:json)?\s*([\s\S]*?)```/i);
+  const body = fenced ? fenced[1]! : raw;
+  const start = body.indexOf("[");
+  const end = body.lastIndexOf("]");
+  if (start < 0 || end <= start) return [];
+  try {
+    const arr = JSON.parse(body.slice(start, end + 1)) as unknown;
+    if (!Array.isArray(arr)) return [];
+    return arr
+      .filter((d): d is Record<string, unknown> => typeof d === "object" && d !== null)
+      .map((d) => ({
+        directive: typeof d.directive === "string" ? d.directive.trim() : "",
+        access: ["read", "write", "full"].includes(String(d.access)) ? String(d.access) : "write",
+      }))
+      .filter((d) => d.directive !== "");
+  } catch {
+    return [];
   }
-  if (task.includes("frontend") || task.includes("ui") || task.includes("client")) {
-    directives.push({ directive: `Implement the frontend/UI components for: ${task}`, access: "write" });
-  }
-  if (task.includes("test") || task.includes("unit")) {
-    directives.push({ directive: `Write unit tests for: ${task}`, access: "read" });
-  }
-  if (directives.length === 0) {
-    for (let i = 0; i < maxParallel; i++) {
-      directives.push({ directive: `Part ${i + 1} of parallel task: ${task}`, access: "write" });
-    }
-  }
-  return directives.slice(0, maxParallel);
 }

src/features/agent/infrastructure/tools/registry.test.ts

@@ -37,12 +37,12 @@ describe("allTools registry", () => {
 
 describe("read-only tools are parallel-safe", () => {
   test("parallel-safe list", () => {
-    for (const name of ["read", "grep", "glob", "semantic_search", "list_symbols", "web_search", "recall", "dir_list", "pong", "seq_think", "dir_cache_update"]) {
+    for (const name of ["read", "grep", "glob", "semantic_search", "list_symbols", "web_search", "recall", "dir_list", "pong", "seq_think"]) {
       expect(toolIsParallelSafe(name)).toBe(true);
     }
   });
-  test("mutating tools are not parallel-safe", () => {
+  test("mutating tools (incl. dir_cache_update, which writes the cache) are not parallel-safe", () => {
-    for (const name of ["edit", "write", "delete", "bash", "git_operator", "git_worktree", "remember", "forget", "todowrite", "plan_enter", "spawn_agents"]) {
+    for (const name of ["edit", "write", "delete", "bash", "git_operator", "git_worktree", "remember", "forget", "todowrite", "plan_enter", "spawn_agents", "dir_cache_update"]) {
       expect(toolIsParallelSafe(name)).toBe(false);
     }
   });

src/features/agent/infrastructure/tools/registry.ts

@@ -88,7 +88,6 @@ export function toolIsParallelSafe(name: string): boolean {
     "dir_list",
     "pong",
     "seq_think",
-    "dir_cache_update",
   ].includes(name);
 }
 

src/features/workflow/infrastructure/complexity.ts

@@ -1,22 +0,0 @@
-/**
- * Complexity heuristics — determine whether a request is complex enough to
- * warrant hive-mind orchestration.
- * Mirrors `apps/infrastructure/src/workflow/hive_mind/complexity.rs`.
- */
-
-/** Heuristics to determine if a request is complex enough for hive-mind. */
-export function isComplexRequest(task: string): boolean {
-  const complexityIndicators = [
-    "refactor",
-    "redesign",
-    "multiple files",
-    "architecture",
-    "migration",
-    "comprehensive",
-    "end-to-end",
-    "full-stack",
-  ];
-
-  const taskLower = task.toLowerCase();
-  return complexityIndicators.some((indicator) => taskLower.includes(indicator));
-}

src/features/workflow/infrastructure/index.ts

@@ -7,5 +7,4 @@ export { executeWorkflowScript, executePrimitive, formatWorkflowResult, MAX_NODE
 export { executeWorkflow } from "./orchestrate.ts";
 export { executeCycle } from "./cycle.ts";
 export { synthesizeConsensus } from "./synthesis.ts";
-export { isComplexRequest } from "./complexity.ts";
 export { writeHiveMindConvergence } from "./docs.ts";

src/features/workflow/infrastructure/script.test.ts

@@ -0,0 +1,61 @@
+/**
+ * Tests for the workflow YAML-subset parser, including the folded multi-line
+ * directive support.
+ */
+import { describe, expect, test } from "bun:test";
+import { parseWorkflowScript } from "./script.ts";
+
+describe("parseWorkflowScript", () => {
+  test("parses a flat single-line workflow", () => {
+    const script = parseWorkflowScript(`
+name: my-workflow
+phases:
+  - name: research
+    directive: "Explore the codebase for the auth module."
+  - name: implement
+    directive: "Implement the change."
+`);
+    expect(script.name).toBe("my-workflow");
+    expect(script.phases.length).toBe(2);
+    expect(script.phases[0]!.name).toBe("research");
+    expect(script.phases[0]!.directive).toBe("Explore the codebase for the auth module.");
+  });
+
+  test("folds an indented multi-line directive into one string", () => {
+    const script = parseWorkflowScript(`
+name: audit
+phases:
+  - name: analyze
+    directive: |
+      Look at two things:
+      first, how auth is wired;
+      second, where secrets are stored.
+`);
+    expect(script.phases.length).toBe(1);
+    const d = script.phases[0]!.directive;
+    expect(d).toContain("Look at two things:");
+    expect(d).toContain("second, where secrets are stored.");
+    // Line breaks are preserved as folded content, not flattened.
+    expect(d).toContain("\n");
+  });
+
+  test("an indented continuation does not swallow the next phase", () => {
+    const script = parseWorkflowScript(`
+phases:
+  - name: a
+    directive: first
+      and more
+  - name: b
+    directive: second
+`);
+    expect(script.phases.length).toBe(2);
+    expect(script.phases[0]!.directive).toContain("and more");
+    expect(script.phases[1]!.name).toBe("b");
+    expect(script.phases[1]!.directive).toBe("second");
+  });
+
+  test("empty or unparseable yaml yields no phases", () => {
+    expect(parseWorkflowScript("").phases.length).toBe(0);
+    expect(parseWorkflowScript("name: x\n").phases.length).toBe(0);
+  });
+});

src/features/workflow/infrastructure/script.ts

@@ -31,21 +31,62 @@ export function parseWorkflowScript(yaml: string): WorkflowScript {
 
   let inPhases = false;
   let currentPhase: Partial<WorkflowPhase> | null = null;
+  /** Set when `directive:` was opened, to fold more-indented lines into it. */
+  let directiveIndent: number | null = null;
+
+  const flushPhase = (): void => {
+    if (currentPhase && (currentPhase.name || currentPhase.directive)) {
+      phases.push({
+        name: currentPhase.name ?? "phase",
+        directive: currentPhase.directive ?? "",
+      });
+    }
+    currentPhase = null;
+    directiveIndent = null;
+  };
+
+  const indentOf = (s: string): number => {
+    const m = s.match(/^(\s*)/);
+    return m ? m[1]!.length : 0;
+  };
+
+  /** Strip one surrounding pair of single/double quotes, YAML-scalar style. */
+  const unquote = (v: string): string => {
+    const t = v.trim();
+    if (t.length >= 2 && ((t.startsWith('"') && t.endsWith('"')) || (t.startsWith("'") && t.endsWith("'")))) {
+      return t.slice(1, -1);
+    }
+    return t;
+  };
 
   for (const raw of lines) {
     const trimmed = raw.trim();
     if (trimmed === "" || trimmed.startsWith("#")) continue;
 
+    // Fold continuation lines into the open directive (indented block text).
+    // A continuation is a non-blank line indented deeper than the directive
+    // key, that is not a new list item and not a `key:` assignment.
+    if (currentPhase && directiveIndent !== null) {
+      const indent = indentOf(raw);
+      const isListItem = trimmed.startsWith("- ");
+      const isKv = /^[\w-]+:\s/.test(trimmed);
+      if (indent > directiveIndent && !isListItem && !isKv) {
+        currentPhase.directive = `${currentPhase.directive ?? ""}\n${trimmed}`;
+        continue;
+      }
+    }
+
     // Top-level key: value
     const topMatch = trimmed.match(/^(\w+):\s*(.*)$/);
     if (topMatch && !raw.startsWith(" ") && !raw.startsWith("-")) {
       const key = topMatch[1]!;
-      const val = topMatch[2]!.trim();
+      const val = unquote(topMatch[2]!);
       if (key === "name" && val !== "") {
         name = val;
       } else if (key === "phases") {
         inPhases = true;
       }
+      directiveIndent = null;
       continue;
     }
 
@@ -54,18 +95,17 @@ export function parseWorkflowScript(yaml: string): WorkflowScript {
     // List item: - name: foo  or  - directive: bar
     const listMatch = trimmed.match(/^-\s+(\w+):\s*(.*)$/);
     if (listMatch) {
-      // If there was a previous phase, push it
+      flushPhase();
-      if (currentPhase && (currentPhase.name || currentPhase.directive)) {
-        phases.push({
-          name: currentPhase.name ?? "phase",
-          directive: currentPhase.directive ?? "",
-        });
-      }
       const key = listMatch[1]!;
-      const val = listMatch[2]!.trim();
+      const val = unquote(listMatch[2]!);
       currentPhase = { name: undefined, directive: undefined };
       if (key === "name") currentPhase.name = val;
-      if (key === "directive") currentPhase.directive = val;
+      if (key === "directive") {
+        currentPhase.directive = val;
+        directiveIndent = indentOf(raw);
+      } else {
+        directiveIndent = null;
+      }
       continue;
     }
 
@@ -74,20 +114,19 @@ export function parseWorkflowScript(yaml: string): WorkflowScript {
       const kvMatch = trimmed.match(/^(\w+):\s*(.*)$/);
       if (kvMatch) {
         const key = kvMatch[1]!;
-        const val = kvMatch[2]!.trim();
+        const val = unquote(kvMatch[2]!);
         if (key === "name") currentPhase.name = val;
-        if (key === "directive") currentPhase.directive = val;
+        if (key === "directive") {
+          currentPhase.directive = val;
+          directiveIndent = indentOf(raw);
+        } else {
+          directiveIndent = null;
+        }
       }
     }
   }
 
-  // Push the last phase
+  flushPhase();
-  if (currentPhase && (currentPhase.name || currentPhase.directive)) {
-    phases.push({
-      name: currentPhase.name ?? "phase",
-      directive: currentPhase.directive ?? "",
-    });
-  }
 
   return { name, phases };
 }

src/features/workflow/infrastructure/synthesis.ts

@@ -8,6 +8,7 @@
  */
 import type { NodeOutput } from "@zesdex/workflow";
 import type { ToolCtx } from "../../agent/infrastructure/tools/mod.ts";
+import { consensusSynthesizerPrompt } from "@zesdex/agent";
 import { MAX_NODE_OUTPUT_CHARS } from "./engine.ts";
 
 /**
@@ -33,16 +34,7 @@ export async function synthesizeConsensus(
 
   const systemMsg = {
     role: "system" as const,
-    content: "You are a consensus synthesizer for a multi-agent hive mind. " +
+    content: consensusSynthesizerPrompt(),
-      "Several independent nodes analysed a problem and produced the outputs " +
-      "below. Distill them into ONE coherent consensus report with these sections:\n" +
-      "- AGREEMENTS: points multiple nodes converge on.\n" +
-      "- CONFLICTS: contradictory conclusions, with which node(s) support each side.\n" +
-      "- KEY FINDINGS: the most important, actionable takeaways.\n" +
-      "- RECOMMENDATION: a single recommended next action, or 'no clear consensus' " +
-      "if the outputs are too divergent.\n" +
-      "Be concise and factual. If a node errored, note it and ignore its content.\n" +
-      "Do not invent facts not present in the node outputs.",
   };
 
   const userMsg = {

src/interfaces/cli/compose.ts

@@ -40,6 +40,12 @@ export interface WiredRuntime {
   apiKey: string;
   model: string;
   apiBase?: string;
+  /** Fetch the provider's available model ids (via GET /models). */
+  listModels?: () => Promise<string[]>;
+  /** Optional LLM max_tokens override from settings. */
+  maxTokens?: number;
+  /** Optional LLM temperature override from settings. */
+  temperature?: number;
 }
 
 /**
@@ -76,7 +82,17 @@ export async function runSingleProcess(): Promise<WiredRuntime> {
     toolDefs(allTools()) as never,
   );
 
-  return { store, executor, turnService, apiKey, model, apiBase: baseUrl };
+  return {
+    store,
+    executor,
+    turnService,
+    apiKey,
+    model,
+    apiBase: baseUrl,
+    listModels: () => llmClient.listModels(),
+    maxTokens: settings.max_tokens,
+    temperature: settings.temperature,
+  };
 }
 
 /** Rooted, long-lived event sink that collects events into an array. */
@@ -111,6 +127,8 @@ export function buildTurnParams(
     api_key: runtime.apiKey,
     model: runtime.model,
     api_base: runtime.apiBase,
+    max_tokens: runtime.maxTokens,
+    temperature: runtime.temperature,
   };
 }
 

src/interfaces/tui/action.ts

@@ -16,8 +16,10 @@ import {
   toastInfo,
   markDirty,
   EditorState,
+  type MenuFrame,
 } from "./state.ts";
 import { gitDiffOutput } from "./git.ts";
+import { modelProviderMenu, configMenu } from "./menu.ts";
 
 /** A single well-typed event in the TUI that mutates `AppStateRest`. */
 export type Action =
@@ -46,7 +48,21 @@ export type Action =
   | { tag: "AbortTurn" }
   | { tag: "Compact" }
   | { tag: "ShowDiff" }
-  | { tag: "DiffScroll"; amount: number };
+  | { tag: "DiffScroll"; amount: number }
+  /* Interactive menu stack */
+  | { tag: "OpenMenu"; frame: MenuFrame }
+  | { tag: "MenuUp" }
+  | { tag: "MenuDown" }
+  | { tag: "MenuPageUp" }
+  | { tag: "MenuPageDown" }
+  | { tag: "MenuHome" }
+  | { tag: "MenuEnd" }
+  | { tag: "MenuSelect" }
+  | { tag: "MenuBack" }
+  | { tag: "MenuBackAll" }
+  | { tag: "MenuApply"; index: number }
+  /** Open a named top-level menu (frame built from live state at apply time). */
+  | { tag: "OpenMenuById"; menu: "model" | "config" };
 
 /** Apply an `Action` to `AppStateRest`. `state` is mutated in place. */
 export function applyAction(state: AppStateRest, action: Action): void {
@@ -184,9 +200,83 @@ export function applyAction(state: AppStateRest, action: Action): void {
         state.dirty = true;
       }
       break;
+
+    case "OpenMenu":
+      state.misc.menuStack.push(action.frame);
+      state.misc.overlay = "menu";
+      state.dirty = true;
+      break;
+    case "MenuUp":
+      moveMenu(state, -1);
+      break;
+    case "MenuDown":
+      moveMenu(state, 1);
+      break;
+    case "MenuPageUp":
+      moveMenu(state, -10);
+      break;
+    case "MenuPageDown":
+      moveMenu(state, 10);
+      break;
+    case "MenuHome": {
+      const f = state.misc.menuStack[state.misc.menuStack.length - 1];
+      if (f && f.items.length) f.selected = 0;
+      state.dirty = true;
+      break;
+    }
+    case "MenuEnd": {
+      const f = state.misc.menuStack[state.misc.menuStack.length - 1];
+      if (f && f.items.length) f.selected = f.items.length - 1;
+      state.dirty = true;
+      break;
+    }
+    case "MenuSelect":
+      // Push nothing here; the controller resolves a sub-menu or action. This
+      // simply marks dirty so the selection is reflected.
+      state.dirty = true;
+      break;
+    case "MenuBack":
+      state.misc.menuStack.pop();
+      if (state.misc.menuStack.length === 0) state.misc.overlay = "none";
+      state.dirty = true;
+      break;
+    case "MenuBackAll":
+      state.misc.menuStack = [];
+      state.misc.overlay = "none";
+      state.dirty = true;
+      break;
+    case "MenuApply":
+      applyMenuAt(state, action.index);
+      break;
+
+    case "OpenMenuById": {
+      const frame =
+        action.menu === "model" ? modelProviderMenu(state) : configMenu(state);
+      state.misc.menuStack.push(frame);
+      state.misc.overlay = "menu";
+      state.dirty = true;
+      break;
+    }
   }
 }
 
+/** Move the top menu's selection by `delta`, wrapping at the ends. */
+function moveMenu(state: AppStateRest, delta: number): void {
+  const frame = state.misc.menuStack[state.misc.menuStack.length - 1];
+  if (!frame || frame.items.length === 0) return;
+  const n = frame.items.length;
+  frame.selected = (frame.selected + delta) % n;
+  if (frame.selected < 0) frame.selected += n;
+  state.dirty = true;
+}
+
+function applyMenuAt(state: AppStateRest, index: number): void {
+  const frame = state.misc.menuStack[state.misc.menuStack.length - 1];
+  if (!frame) return;
+  state.misc.selectedIndex = index;
+  state.dirty = true;
+}
+
 /** Periodic timer tick — drains queued events and applies streaming text. */
 function handleTick(state: AppStateRest): void {
   // Streaming updates are applied via appendToLastTranscript by the turn

src/interfaces/tui/command.ts

@@ -14,8 +14,8 @@ export type Command =
   | { tag: "Clear" }
   | { tag: "ClearConfirm" }
   | { tag: "Edit"; path: string }
-  | { tag: "McpAdd"; name: string; command: string }
+  | { tag: "OpenModelMenu" }
-  | { tag: "ModelList" }
+  | { tag: "OpenConfigMenu" }
   | { tag: "Compact" }
   | { tag: "TodoOpen" }
   | { tag: "PlanOpen" }
@@ -28,10 +28,9 @@ export function parseCommand(text: string): Command {
   const trimmed = text.trim();
   if (!trimmed.startsWith("/")) return { tag: "Unknown", text: trimmed };
 
-  const parts = trimmed.split(" ", 3);
+  const parts = trimmed.split(" ", 2);
   const cmd = parts[0] ?? "";
   const arg1 = parts[1] ?? "";
-  const arg2 = parts[2] ?? "";
 
   switch (cmd) {
     case "/help":
@@ -43,18 +42,15 @@ export function parseCommand(text: string): Command {
     case "/edit":
       return { tag: "Edit", path: arg1 === "" ? "." : arg1 };
     case "/mcp":
+      // No inline sub-commands: `/mcp` alone opens the interactive server list.
       if (arg1 === "") return { tag: "McpOpen" };
-      if (arg1 === "add") {
-        const rest = arg2.trim();
-        const space = rest.indexOf(" ");
-        if (space !== -1) {
-          return { tag: "McpAdd", name: rest.slice(0, space), command: rest.slice(space + 1).trim() };
-        }
-        return { tag: "McpAdd", name: rest, command: "" };
-      }
       return { tag: "Unknown", text: cmd };
     case "/model":
-      return { tag: "ModelList" };
+      // No inline `/model add` / `/model ls`: `/model` is a nested picker.
+      return { tag: "OpenModelMenu" };
+    case "/config":
+    case "/settings":
+      return { tag: "OpenConfigMenu" };
     case "/compact":
       return { tag: "Compact" };
     case "/todo":
@@ -85,10 +81,10 @@ export function applyCommand(cmd: Command): Action[] {
       return [{ tag: "OpenOverlay", overlay: "clear_confirm" }];
     case "Edit":
       return [{ tag: "OpenEditor", path: cmd.path }];
-    case "McpAdd":
+    case "OpenModelMenu":
-      return [{ tag: "McpAdd", name: cmd.name, command: cmd.command }];
+      return [{ tag: "OpenMenuById", menu: "model" }];
-    case "ModelList":
+    case "OpenConfigMenu":
-      return [{ tag: "ModelList" }];
+      return [{ tag: "OpenMenuById", menu: "config" }];
     case "Compact":
       return [{ tag: "Compact" }];
     case "TodoOpen":

src/interfaces/tui/controller.ts

@@ -7,9 +7,10 @@
  */
 import { Roles } from "@zesdex/domain";
 import type { AppStateRest } from "./state.ts";
-import { AutocompleteKind, markDirty, toastInfo, overlayIsActive } from "./state.ts";
+import { AutocompleteKind, COMMANDS, markDirty, toastInfo, overlayIsActive } from "./state.ts";
 import type { Action } from "./action.ts";
 import { applyCommand, parseCommand } from "./command.ts";
+import { resolveMenu } from "./menu.ts";
 
 /** A decoded terminal key event (normalised from raw input). */
 export interface KeyEvent {
@@ -32,6 +33,8 @@ export const KeyCode = {
   Down: "ArrowDown",
   PageUp: "PageUp",
   PageDown: "PageDown",
+  Home: "Home",
+  End: "End",
   Esc: "Escape",
   Tab: "Tab",
   Char: "Char",
@@ -183,7 +186,48 @@ export function handleKey(key: KeyEvent, state: AppStateRest): Action[] {
     return [];
   }
 
+  // ── Interactive menu stack ───────────────────────────────────────────
+  if (state.misc.menuStack.length > 0) {
+    const frame = state.misc.menuStack[state.misc.menuStack.length - 1];
+    const n = frame?.items.length ?? 0;
+    if (code === KeyCode.Char && key.ctrl && key.value === "c") return [{ tag: "QuitConfirm" }];
+    if (code === KeyCode.Esc) return [{ tag: "MenuBack" }];
+    if (code === KeyCode.Up) return n > 0 ? [{ tag: "MenuUp" }] : [];
+    if (code === KeyCode.Down) return n > 0 ? [{ tag: "MenuDown" }] : [];
+    if (code === KeyCode.PageUp) return n > 0 ? [{ tag: "MenuPageUp" }] : [];
+    if (code === KeyCode.PageDown) return n > 0 ? [{ tag: "MenuPageDown" }] : [];
+    if (code === KeyCode.Home) return n > 0 ? [{ tag: "MenuHome" }] : [];
+    if (code === KeyCode.End) return n > 0 ? [{ tag: "MenuEnd" }] : [];
+    if (code === KeyCode.Enter) {
+      return enterMenu(state, frame?.id ?? "");
+    }
+    // vim-style j/k / g/G navigation.
+    if (code === KeyCode.Char && key.value === "j") return n > 0 ? [{ tag: "MenuDown" }] : [];
+    if (code === KeyCode.Char && key.value === "k") return n > 0 ? [{ tag: "MenuUp" }] : [];
+    if (code === KeyCode.Char && key.value === "g") return n > 0 ? [{ tag: "MenuHome" }] : [];
+    if (code === KeyCode.Char && key.value === "G") return n > 0 ? [{ tag: "MenuEnd" }] : [];
+    // Backspace removes the last filter character (type-to-filter).
+    if (code === KeyCode.Backspace) return menuFilterBackspace(state);
+    // Printable char → type-to-filter (narrow + jump to first match).
+    if (code === KeyCode.Char && key.value && key.value.trim().length > 0 && /^[\w./-]$/.test(key.value)) {
+      return menuFilterAppend(state, key.value);
+    }
+    return [];
+  }
+
   // ── Normal (non-overlay) dispatch ────────────────────────────────────
+  // Slash commands can run even when an overlay is open, BUT an open
+  // autocomplete dropdown has first claim on Enter: it completes the command
+  // token rather than firing a half-typed "/mo" as an unknown command.
+  if (
+    code === KeyCode.Enter &&
+    state.input.buffer.startsWith("/") &&
+    !state.input.autocompleteVisible
+  ) {
+    const text = state.input.buffer;
+    return applyCommand(parseCommand(text));
+  }
+
   switch (code) {
     case KeyCode.Char:
       if (key.ctrl && key.value === "c") return [{ tag: "QuitConfirm" }];
@@ -215,7 +259,17 @@ export function handleKey(key: KeyEvent, state: AppStateRest): Action[] {
       return [];
     case KeyCode.Enter:
       if (state.input.autocompleteVisible) {
-        state.input.selectAutocomplete();
+        const wasCommand = state.input.autocompleteKind === AutocompleteKind.Command;
+        const selected = state.input.selectAutocomplete();
+        // Completing a command via Enter should also RUN it (single Enter),
+        // unless the user typed more args after the token (e.g. "/edit foo").
+        if (selected && wasCommand) {
+          const line = state.input.buffer.trim();
+          const word = line.split(/\s+/)[0] ?? "";
+          if (COMMANDS.includes(word)) {
+            return applyCommand(parseCommand(line));
+          }
+        }
         return mark(state);
       }
       if (overlayIsActive(state.misc.overlay)) {
@@ -223,9 +277,6 @@ export function handleKey(key: KeyEvent, state: AppStateRest): Action[] {
       }
       {
         const text = state.input.buffer;
-        if (text.startsWith("/")) {
-          return applyCommand(parseCommand(text));
-        }
         return [{ tag: "SubmitInput", text }];
       }
     case KeyCode.Backspace:
@@ -249,14 +300,6 @@ export function handleKey(key: KeyEvent, state: AppStateRest): Action[] {
         state.input.cycleAutocomplete(false);
         mark(state);
         return [];
-      } else if (state.misc.overlay === "model_selector") {
-        state.misc.selectedIndex = cycleSelectedIndex(
-          state.misc.selectedIndex,
-          Object.keys(state.appConfig.providers).length,
-          false,
-        );
-        mark(state);
-        return [];
       } else if (key.ctrl) {
         return [{ tag: "ScrollUp" }];
       } else {
@@ -267,14 +310,6 @@ export function handleKey(key: KeyEvent, state: AppStateRest): Action[] {
         state.input.cycleAutocomplete(true);
         mark(state);
         return [];
-      } else if (state.misc.overlay === "model_selector") {
-        state.misc.selectedIndex = cycleSelectedIndex(
-          state.misc.selectedIndex,
-          Object.keys(state.appConfig.providers).length,
-          true,
-        );
-        mark(state);
-        return [];
       } else if (key.ctrl) {
         return [{ tag: "ScrollDown" }];
       } else {
@@ -315,6 +350,50 @@ export function handleKey(key: KeyEvent, state: AppStateRest): Action[] {
   }
 }
 
+/**
+ * Resolve Enter on a menu item. `id` joins the menu navigation path with `:`
+ * (e.g. `model:provider`, `model:model`, `settings:internet`). Returns the
+ * actions to run — typically an `OpenMenu` for a sub-menu, or an apply action
+ * that pops the whole stack.
+ */
+function enterMenu(state: AppStateRest, id: string): Action[] {
+  // Delegate Enter to the semantic resolver (menu.ts): it returns an OpenMenu
+  // for nested pickers, or applied actions that pop the stack.
+  return resolveMenu(state, id) ?? [];
+}
+
+/** Append a character to the type-to-filter and jump to the first match. */
+function menuFilterAppend(state: AppStateRest, ch: string): Action[] {
+  const frame = state.misc.menuStack[state.misc.menuStack.length - 1];
+  if (!frame || frame.items.length === 0) return [];
+  frame.filter = (frame.filter ?? "") + ch.toLowerCase();
+  jumpFiltered(frame);
+  markDirty(state);
+  return [];
+}
+
+/** Remove the last filter character; jump back to the first match. */
+function menuFilterBackspace(state: AppStateRest): Action[] {
+  const frame = state.misc.menuStack[state.misc.menuStack.length - 1];
+  if (!frame || frame.items.length === 0) return [];
+  const f = frame.filter ?? "";
+  frame.filter = f.slice(0, -1);
+  jumpFiltered(frame);
+  markDirty(state);
+  return [];
+}
+
+/** Move selection to the first item matching the frame filter (if any). */
+function jumpFiltered(frame: { items: string[]; selected: number; filter?: string }): void {
+  const f = (frame.filter ?? "").toLowerCase();
+  if (!f) {
+    frame.selected = 0;
+    return;
+  }
+  const hit = frame.items.findIndex((it) => it.toLowerCase().includes(f));
+  if (hit !== -1) frame.selected = hit;
+}
+
 /** Local helper: whether an agent turn is in flight. */
 function turnInFlightLocal(state: AppStateRest): boolean {
   return state.turnInFlightFlag.value;

src/interfaces/tui/menu.ts

@@ -0,0 +1,160 @@
+/**
+ * Data-driven interactive menus for the TUI.
+ *
+ * Replaces inline sub-commands (`/model add`, `/mcp add …`) with pickers the
+ * way Claude Code / opencode do. Each menu is a flat `MenuFrame` of selectable
+ * rows; choosing a row either opens a nested sub-menu (via `OpenMenu`) or
+ * applies a change and pops the stack. Everything here is logic-pure and
+ * unit-testable.
+ *
+ * Every select dialog shares one keymap (arrow/jk, Home/End, PageUp/Down,
+ * Enter select, Esc back) centralised in `controller.ts`; the *shape* of each
+ * menu and the *effect* of choosing a row lives here. The current value is
+ * shown on every relevant row.
+ */
+import type { AppStateRest, MenuFrame } from "./state.ts";
+import type { Action } from "./action.ts";
+import { toastInfo } from "./state.ts";
+
+/* ── /model tree ──────────────────────────────────────────────────── */
+
+/** Build the `/model` provider menu (opencode-style provider list first). */
+export function modelProviderMenu(state: AppStateRest): MenuFrame {
+  const names = Object.keys(state.appConfig.providers);
+  const list = names.length ? names : ["claude", "zen", "router"];
+  const current = String(state.settings.provider ?? "");
+  return {
+    id: "model:provider",
+    title: "Providers",
+    items: list,
+    details: list.map((p) =>
+      p === current ? `   (active)` : "",
+    ),
+    hint: "↑/↓ or j/k · Enter open models · Esc back",
+    selected: Math.max(0, list.indexOf(current)),
+  };
+}
+
+/** Build the per-provider model sub-menu, preferring models fetched live. */
+export function modelSubMenu(state: AppStateRest, provider: string): MenuFrame | null {
+  const live = state.misc.availableModels;
+  const current = String(state.settings.model ?? "");
+  if (live.length > 0) {
+    return {
+      id: `model:${provider}`,
+      title: `Models — ${provider} (${live.length})`,
+      items: [...live],
+      details: live.map((m) => (m === current ? "  (current)" : (m === state.settings.model ? "  (current)" : ""))),
+      hint: "Enter applies · Esc back to providers",
+      selected: Math.max(0, live.indexOf(current)),
+    };
+  }
+  // No /models endpoint (or it's still fetching) — degrade to a small static
+  // catalog so the picker never blocks.
+  const known = ["claude-opus-5", "claude-sonnet-5", "claude-haiku-4-5", "deepseek-v4-flash-free", "gpt-5"];
+  return {
+    id: `model:${provider}`,
+    title: `Models — ${provider}`,
+    items: known,
+    details: known.map((m) => (m === current ? "  (current)" : "")),
+    hint: "live /models unavailable — fallback list · Enter applies",
+    selected: Math.max(0, known.indexOf(current)),
+  };
+}
+
+/* ── /config tree ─────────────────────────────────────────────────── */
+
+/** Build the top-level `/config` menu. */
+export function configMenu(state: AppStateRest): MenuFrame {
+  const internet = String(state.settings.internet_mode ?? "Off");
+  const review = state.settings.review_enabled !== false ? "on" : "off";
+  const provider = String(state.settings.provider ?? "");
+  return {
+    id: "config:root",
+    title: "Settings",
+    items: ["Internet access", "Review enabled", "Provider / model"],
+    details: [
+      `current: ${internet}`,
+      `current: ${review}`,
+      `current: ${provider}`,
+    ],
+    hint: "↑/↓ or j/k · Enter · Esc close",
+    selected: 0,
+  };
+}
+
+/** Build the internet-mode sub-menu. */
+function internetMenu(state: AppStateRest): MenuFrame {
+  const modes = ["Off", "ReadOnly", "Full"];
+  const current = String(state.settings.internet_mode ?? "Off");
+  return {
+    id: "config:internet",
+    title: "Config — internet access",
+    items: modes,
+    details: modes.map((m) => (m === current ? "  (current)" : "")),
+    hint: "Enter set · Esc back",
+    selected: Math.max(0, modes.indexOf(current)),
+  };
+}
+
+/* ── Menu resolver ────────────────────────────────────────────────── */
+
+/**
+ * Resolve Enter inside any interactive menu. `id` encodes the navigation path
+ * (e.g. `model:provider`, `model:zen`, `config:root`, `config:internet`).
+ * Reads the top-most frame's selected row and returns the actions to run —
+ * usually an `OpenMenu` for a nested picker, or applied actions that pop the
+ * entire stack.
+ */
+export function resolveMenu(state: AppStateRest, id: string): Action[] {
+  const frame = state.misc.menuStack[state.misc.menuStack.length - 1];
+  if (!frame) return [];
+  const item = frame.items[frame.selected];
+  if (item === undefined) return [];
+  const idx = frame.selected;
+
+  // ── model tree ────────────────────────────────────────────────────
+  if (id === "model:provider") {
+    const sub = modelSubMenu(state, item);
+    return sub ? [{ tag: "OpenMenu", frame: sub }] : popAll();
+  }
+  if (id.startsWith("model:")) {
+    // A model row under a provider — apply it.
+    const provider = id.slice("model:".length);
+    state.settings = { ...state.settings, provider, model: item };
+    toastInfo(state, `Using ${provider} / ${item}`);
+    return popAll();
+  }
+
+  // ── config tree ───────────────────────────────────────────────────
+  if (id === "config:root") {
+    switch (idx) {
+      case 0:
+        return [{ tag: "OpenMenu", frame: internetMenu(state) }];
+      case 1: {
+        const on = state.settings.review_enabled !== false;
+        state.settings = { ...state.settings, review_enabled: !on };
+        toastInfo(state, `Review ${on ? "disabled" : "enabled"}`);
+        return popAll();
+      }
+      case 2:
+        return [{ tag: "OpenMenu", frame: modelProviderMenu(state) }];
+      default:
+        return [];
+    }
+  }
+  if (id === "config:internet") {
+    const modes = ["Off", "ReadOnly", "Full"];
+    const chosen = modes[idx] ?? "Off";
+    state.settings = { ...state.settings, internet_mode: chosen };
+    toastInfo(state, `Internet mode → ${chosen}`);
+    return popAll();
+  }
+
+  return [];
+}
+
+/** Pop the whole menu stack (a leaf decision applied). */
+function popAll(): Action[] {
+  return [{ tag: "MenuBackAll" }];
+}

src/interfaces/tui/render.smoke.test.tsx

@@ -0,0 +1,160 @@
+/**
+ * Smoke-render harness for the new TUI layout. Renders `ZesdexApp` through the
+ * real OpenTUI test renderer across a few representative states and snapshots
+ * the characters so we can eyeball the design without a live terminal.
+ */
+import { describe, test, expect } from "bun:test";
+import { createTestRenderer } from "@opentui/core/testing";
+import { createRoot, flushSync } from "@opentui/react";
+import { ZesdexApp } from "./ui.tsx";
+import { applyAction } from "./action.ts";
+import {
+  createTuiState,
+  pushTranscript,
+  makeChatMessage,
+  type AppStateRest,
+} from "./state.ts";
+import { Roles } from "@zesdex/domain";
+
+async function render(state: AppStateRest, w = 80, h = 24): Promise<string> {
+  const setup = await createTestRenderer({ width: w, height: h });
+  const root = createRoot(setup.renderer);
+  (globalThis as unknown as Record<string, unknown>).IS_REACT_ACT_ENVIRONMENT = true;
+  flushSync(() => root.render(<ZesdexApp state={state} onQuit={() => {}} />));
+  await setup.renderOnce();
+  await setup.flush();
+  const frame = setup.captureCharFrame();
+  flushSync(() => root.unmount());
+  await setup.renderOnce();
+  await setup.flush();
+  (globalThis as unknown as Record<string, unknown>).IS_REACT_ACT_ENVIRONMENT = false;
+  return frame;
+}
+
+describe("ZesdexApp render smoke", () => {
+  test("empty transcript shows prompt + placeholder", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    const frame = await render(s);
+    expect(frame).toContain("zesdex");
+    expect(frame).toContain("Ask Zesdex to help");
+    expect(frame).toContain("● idle");
+  });
+
+  test("with instructions renders the folder name label", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    await render(s);
+    expect(1).toBe(1);
+  });
+
+  test("user + assistant markdown message renders", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    pushTranscript(s, makeChatMessage(Roles.User, "hello there"));
+    pushTranscript(s, makeChatMessage(Roles.Assistant, "# Result\n\nSome **bold** and `code`."));
+    const frame = await render(s);
+    expect(frame).toContain("[YOU]");
+    expect(frame).toContain("[AI]");
+    expect(frame).toContain("Result");
+  });
+
+  test("help overlay renders the shortcut text", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.misc.overlay = "help";
+    const frame = await render(s);
+    expect(frame).toContain("Keyboard Shortcuts");
+    expect(frame).toContain("Ctrl+C");
+  });
+
+  test("model selector overlay renders provider list", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.misc.overlay = "model_selector";
+    s.appConfig = { providers: { claude: {}, zen: {}, router: {} } };
+    s.settings = { provider: "" }; // no active provider → none marked (active)
+    const frame = await render(s, 80, 24);
+    // Count how many distinct provider names actually hit the frame.
+    const count = ["claude","zen","router"].filter((p) => frame.includes(p)).length;
+    console.log("MODEL-DEBUG providers-in-frame:", count, frame.split("\n").filter((l)=>l.includes("▸")).map((l)=>l.trim()));
+    expect(count).toBe(3);
+  });
+
+  test("running state shows spinner + activity + token usage", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.turnInFlightFlag.value = true;
+    s.misc.activity = [
+      { label: "grep", kind: "tool" },
+      { label: "read", kind: "tool", count: 3 },
+    ];
+    s.lastUsage = { tokensIn: 12, tokensOut: 34 };
+    const frame = await render(s);
+    expect(frame).toContain("● running");
+    expect(frame).toContain("grep");
+    expect(frame).toContain("read×3");
+    expect(frame).toContain("in:12");
+    expect(frame).toContain("out:34");
+    expect(frame).toContain("Esc abort");
+  });
+
+  test("diff overlay renders git diff lines with colour mapping", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.misc.overlay = "diff";
+    s.misc.diffContent = "diff --git a/x b/x\n@@ -1,2 +1,3 @@\n removed\n+added\n context\n";
+    const frame = await render(s, 90, 24);
+    expect(frame).toContain("Git Diff");
+    expect(frame).toContain("diff --git");
+    expect(frame).toContain("removed");
+    expect(frame).toContain("+added");
+    expect(frame).toContain("context");
+  });
+
+  test("editor overlay renders file buffer with line numbers", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.misc.overlay = "editor";
+    s.misc.editor = {
+      path: "src/a.ts",
+      content: "line one\nline two\nline three",
+      cursor: 3,
+      asString: () => "",
+      deleteLeft: () => {},
+    } as any;
+    const frame = await render(s, 80, 24);
+    expect(frame).toContain("src/a.ts");
+    expect(frame).toContain("line one");
+    expect(frame).toContain("line three");
+  });
+
+  test("interactive /model menu renders providers with active preselected", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.appConfig = { providers: { claude: {}, zen: {}, router: {} } };
+    s.settings = { provider: "zen" };
+    applyAction(s, { tag: "OpenMenuById", menu: "model" });
+    const frame = await render(s, 90, 24);
+    expect(frame).toContain("Providers");
+    expect(frame).toContain("claude");
+    expect(frame).toContain("zen");
+    expect(frame).toContain("router");
+    expect(frame).toContain("type to filter");
+    expect(frame).toContain("(active)");
+  });
+
+  test("autocomplete dropdown renders command rows (not overlapping)", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.input.buffer = "/mo";
+    s.input.cursor = 3;
+    s.input.openAutocomplete();
+    const frame = await render(s, 80, 20);
+    expect(frame).toContain("commands");
+    expect(frame).toContain("/model");
+    expect(frame).toContain("Tab accept");
+  });
+
+  test("interactive /config menu shows rows with current values", async () => {
+    const s = createTuiState(["/tmp/x"], "/tmp/x/s", "/tmp/x/m");
+    s.settings = { provider: "zen", internet_mode: "Off", review_enabled: true };
+    applyAction(s, { tag: "OpenMenuById", menu: "config" });
+    const frame = await render(s, 90, 24);
+    expect(frame).toContain("Settings");
+    expect(frame).toContain("Internet access");
+    expect(frame).toContain("Review enabled");
+    expect(frame).toContain("current: Off");
+    expect(frame).toContain("current: zen");
+  });
+});

src/interfaces/tui/run.ts

@@ -54,10 +54,31 @@ function makeTurnRunner(runtime: WiredRuntime, state: AppStateRest): TurnRunner
       api_key: runtime.apiKey,
       model: runtime.model,
       api_base: runtime.apiBase,
+      max_tokens: runtime.maxTokens,
+      temperature: runtime.temperature,
     });
   };
 }
 
+/** Append an entry to the bounded live-activity breadcrumb. */
+function addActivity(
+  state: AppStateRest,
+  entry: { label: string; detail?: string; kind: "tool" | "agent" | "error" },
+): void {
+  const activity = state.misc.activity;
+  // Dedupe consecutive repeats of the same tool so `read read read` collapses.
+  const last = activity[activity.length - 1];
+  const key = `${entry.label}:${entry.kind}`;
+  if (last && `${last.label}:${last.kind}` === key && last.detail === entry.detail) {
+    last.count = (last.count ?? 1) + 1;
+  } else {
+    activity.push({ ...entry, count: 1 });
+  }
+  // Keep the breadcrumb tight (max 12 entries).
+  while (activity.length > 12) activity.shift();
+  state.dirty = true;
+}
+
 /** Route one live `TurnEvent` into TUI state (transcript, status bar, toasts). */
 function handleEvent(state: AppStateRest, ev: TurnEvent): void {
   switch (ev.kind) {
@@ -74,38 +95,34 @@ function handleEvent(state: AppStateRest, ev: TurnEvent): void {
       );
       break;
     case "tool_result":
-      pushTranscript(
+      // A tool result is a breadcrumb, not a full transcript entry — keep the
-        state,
+      // transcript clean and fold the tool name into the live activity bar.
-        makeChatMessage(
+      addActivity(state, {
-          Roles.Tool,
+        label: ev.tool_name,
-          `→ ${ev.tool_name}${ev.is_error ? " (error)" : ""}`,
+        detail: ev.is_error ? "error" : undefined,
-        ),
+        kind: ev.is_error ? "error" : "tool",
-      );
+      });
       break;
     case "usage":
       state.lastUsage = { tokensIn: ev.tokens_in, tokensOut: ev.tokens_out };
       state.dirty = true;
       break;
     case "agent_progress":
-      pushTranscript(
+      addActivity(state, {
-        state,
+        label: ev.progress.agent_name,
-        makeChatMessage(
+        detail: ev.progress.error ?? ev.progress.current_tool ?? undefined,
-          Roles.Tool,
+        kind: ev.progress.error ? "error" : "agent",
-          `[${ev.progress.agent_name}] ${ev.progress.status}${ev.progress.error ? ` — ${ev.progress.error}` : ev.progress.current_tool ? ` · ${ev.progress.current_tool}` : ""}`,
+      });
-        ),
-      );
       break;
     case "workflow_agent_update":
-      pushTranscript(
+      addActivity(state, {
-        state,
+        label: ev.agent_name,
-        makeChatMessage(
+        detail: ev.error ?? undefined,
-          Roles.Tool,
+        kind: ev.error ? "error" : "agent",
-          `[${ev.agent_name}] ${ev.status}${ev.error ? ` — ${ev.error}` : ""}`,
+      });
-        ),
-      );
       break;
     case "error":
-      pushTranscript(state, makeChatMessage(Roles.System, `error: ${ev.message}`));
+      addActivity(state, { label: "error", detail: ev.message, kind: "error" });
       toastError(state, `Turn error: ${ev.message}`);
       break;
     case "compacted":
@@ -150,5 +167,16 @@ export async function runTui(): Promise<void> {
     sessionId: state.sessionId,
   } as Record<string, unknown>;
 
+  // Prime the /models list in the background so /model shows live models
+  // instead of a hardcoded catalog; the menu falls back until it lands.
+  if (runtime.listModels) {
+    void runtime.listModels().then((models) => {
+      if (models.length > 0) {
+        state.misc.availableModels = models;
+        state.dirty = true;
+      }
+    });
+  }
+
   await bootTui(state);
 }

src/interfaces/tui/state.ts

@@ -36,13 +36,14 @@ export const COMMANDS: string[] = [
   "/quit",
   "/clear",
   "/edit",
-  "/mcp add",
+  "/mcp",
   "/model",
-  "/model ls",
+  "/config",
-  "/model add",
   "/todo",
   "/usage",
   "/compact",
+  "/diff",
+  "/plan",
 ];
 
 /** The user's input buffer, cursor, history, and autocomplete state. */
@@ -56,6 +57,8 @@ export class InputState {
   autocompleteVisible = false;
   autocompleteKind: AutocompleteKind = AutocompleteKind.Command;
   mentionStart = 0;
+  /** Start index of the command token being completed (for in-place replace). */
+  autocompleteTokenStart = 0;
   historyFile: string | null = null;
 
   /** Hide the autocomplete dropdown and clear its state. */
@@ -65,6 +68,7 @@ export class InputState {
     this.autocompleteIdx = 0;
     this.autocompleteKind = AutocompleteKind.Command;
     this.mentionStart = 0;
+    this.autocompleteTokenStart = 0;
   }
 
   /** Open or refresh the autocomplete dropdown by filtering COMMANDS. */
@@ -74,11 +78,28 @@ export class InputState {
       this.closeAutocomplete();
       return;
     }
-    const prefix = trimmed.toLowerCase();
+    // Filter on the *command token* (the word before any space), so typing
-    this.autocompleteCandidates = COMMANDS.filter((c) => c.startsWith(prefix));
+    // "/edit foo" still matches "/edit". A trailing/embedded space must not
+    // kill the dropdown.
+    const word = trimmed.split(/\s+/)[0] ?? "";
+    const prefix = word.toLowerCase();
+    if (!prefix.startsWith("/")) {
+      this.closeAutocomplete();
+      return;
+    }
+    // Prefer exact-match-first so a fully-typed command jumps to the top.
+    this.autocompleteCandidates = COMMANDS.filter((c) => c.startsWith(prefix)).sort(
+      (a, b) => Number(a === word) - Number(b === word),
+    );
+    // Record where the command token starts in the *raw buffer* (not the
+    // trimmed word) so selectAutocomplete can replace just that token.
+    this.autocompleteTokenStart = this.buffer.indexOf("/");
     this.autocompleteKind = AutocompleteKind.Command;
     this.autocompleteIdx = 0;
-    this.autocompleteVisible = this.autocompleteCandidates.length > 0;
+    // A fully-typed command needs no dropdown — entering closes it so Enter
+    // runs the command immediately instead of "selecting" the same string.
+    const exact = this.autocompleteCandidates.includes(word);
+    this.autocompleteVisible = this.autocompleteCandidates.length > 0 && !exact;
   }
 
   /** Find the `@mention` token (if any) immediately before the cursor. */
@@ -129,13 +150,26 @@ export class InputState {
     }
   }
 
+  /** Index just past the command token (stops at a space or end of buffer). */
+  private endOfToken(): number {
+    const start = this.autocompleteTokenStart;
+    if (start < 0 || start > this.buffer.length) return start;
+    const rest = this.buffer.slice(start);
+    const space = rest.search(/\s/);
+    return space === -1 ? this.buffer.length : start + space;
+  }
+
   /** Accept the currently selected autocomplete candidate. Returns true on success. */
   selectAutocomplete(): boolean {
     const candidate = this.autocompleteCandidates[this.autocompleteIdx];
     if (candidate === undefined || candidate === null) return false;
     if (this.autocompleteKind === AutocompleteKind.Command) {
-      this.buffer = candidate;
+      // Replace only the command token (e.g. "/mo" → "/model"), keeping any
-      this.cursor = this.buffer.length;
+      // text the user already typed after it (e.g. "/edit path" stays).
+      const start = this.autocompleteTokenStart;
+      const end = this.endOfToken();
+      this.buffer = this.buffer.slice(0, start) + candidate + this.buffer.slice(end);
+      this.cursor = start + candidate.length;
     } else {
       if (this.cursor < this.mentionStart || this.mentionStart > this.buffer.length) {
         this.closeAutocomplete();
@@ -194,6 +228,12 @@ export class InputState {
     return result;
   }
 
+  /** Clear the input buffer and reset the cursor. */
+  clear(): void {
+    this.buffer = "";
+    this.cursor = 0;
+  }
+
   /** Navigate backward through input history. */
   historyUp(): void {
     if (this.history.length === 0) return;
@@ -248,6 +288,7 @@ export const Overlay = {
   ModelSelector: "model_selector",
   ClearConfirm: "clear_confirm",
   Diff: "diff",
+  Menu: "menu",
 } as const;
 export type Overlay = (typeof Overlay)[keyof typeof Overlay];
 
@@ -277,6 +318,49 @@ export class EditorState {
   }
 }
 
+/* ── Turn activity ────────────────────────────────────────────────── */
+
+/** A compact, bounded breadcrumb of agent activity during a turn. */
+export interface ActivityEntry {
+  /** Short label, e.g. a tool name or agent name. */
+  label: string;
+  /** Optional detail (current tool, error). */
+  detail?: string;
+  /** Dramatic colouring: "tool" | "agent" | "error". */
+  kind: "tool" | "agent" | "error";
+  /** Number of consecutive repeats (collapsed into one chip). */
+  count?: number;
+}
+
+/* ── Interactive menu ─────────────────────────────────────────────── */
+
+/**
+ * A single frame in the interactive menu stack. Reusable across every
+ * configuration surface (models, settings, MCP, todo, …) so each is a
+ * picker like Claude Code / opencode — never inline sub-commands.
+ */
+export interface MenuFrame {
+  /** Stable id used by the resolver to route Enter (e.g. "model:provider"). */
+  id: string;
+  title: string;
+  /** Selectable option labels (one line each). */
+  items: string[];
+  /** Optional dim detail shown under each option. */
+  details?: string[];
+  /** Footer hint line (keybindings / current value). */
+  hint?: string;
+  selected: number;
+  /** Live type-to-filter: typing narrows/highlights matching rows. */
+  filter?: string;
+}
+
+export const emptyMenuFrame = (): MenuFrame => ({
+  id: "",
+  title: "",
+  items: [],
+  selected: 0,
+});
+
 /* ── Misc state ───────────────────────────────────────────────────── */
 
 /** The "miscellaneous" slice of app state. */
@@ -296,6 +380,12 @@ export class MiscState {
   pendingClipboardCopy: string | null = null;
   diffContent = "";
   diffScroll = 0;
+  /** Breadcrumb of the current turn's activity (tools / sub-agents). */
+  activity: ActivityEntry[] = [];
+  /** Stack of interactive menu frames (top = currently shown). */
+  menuStack: MenuFrame[] = [];
+  /** Model ids fetched live from the provider's /models endpoint. */
+  availableModels: string[] = [];
 
   pushToast(toast: { kind: string; message: string; createdMs: number; ttlMs: number }): void {
     this.toasts.push(toast);
@@ -329,10 +419,13 @@ export interface ChatMessageDisplay {
   content: string;
   reasoning: string;
   timestamp: number;
+  /** Sliding-window replay cursor for content + reasoning (transient). */
+  _posContent: number;
+  _posReasoning: number;
 }
 
 export function makeChatMessage(role: Role, content: string): ChatMessageDisplay {
-  return { role, content, reasoning: "", timestamp: Date.now() };
+  return { role, content, reasoning: "", timestamp: Date.now(), _posContent: 0, _posReasoning: 0 };
 }
 
 /** Bounded ring of recent chat messages used to render the transcript view. */
@@ -357,9 +450,9 @@ export class TranscriptCache {
     const msg = this.messages[this.messages.length - 1];
     if (msg && msg.role === Roles.Assistant) {
       if (isReasoning) {
-        msg.reasoning = reconcileStream(msg.reasoning, text);
+        msg.reasoning = reconcileStream(msg.reasoning, text, msg, true);
       } else {
-        msg.content = reconcileStream(msg.content, text);
+        msg.content = reconcileStream(msg.content, text, msg, false);
       }
       this.dirty = true;
     }
@@ -367,34 +460,72 @@ export class TranscriptCache {
 }
 
 /**
- * Reconcile an incoming stream chunk against the accumulated text.
+ * Reconcile an incoming stream chunk against the accumulated text, deduping the
+ * sliding-window replays that 9router's OpenAI-compatible layer emits.
  *
  * Streaming backends disagree on chunk semantics:
  *  - Native OpenAI/Anthropic send small *incremental* deltas.
- *  - Some gateways/proxies (e.g. 9router's OpenAI-compatible layer) emit
+ *  - 9router/gateways re-send the FULL accumulated text as a repeated window:
- *    *full-text* chunks, where each chunk is the complete transcript so far.
+ *    each "wave" replays the already-emitted segments from the start, then one
+ *    fresh token. So the stream is s1,s2,…,sk then s1,s2,…,sk,s_{k+1} then
+ *    s1,…,s_{k+1},s_{k+2} … where s_i + … = the final text.
  *
- * Blindly concatenating full-text chunks duplicates content (partial + full),
+ * Naive concatenation doubles the text (garbled output). We track a per-stream
- * which manifests as garbled messages. This helper accepts both modes:
+ * replay cursor `pos` into `accumulated`: a chunk that continues at `pos` (or
- *   - a chunk that is a strict prefix of the tail → a shorter replay → ignore;
+ * restarts at 0, i.e. begins a fresh wave) is a replay → ignore it and advance
- *   - a chunk that extends/replays the whole accumulated text → take it as the
+ * the cursor; only a chunk that matches nothing already present is brand-new
- *     new authoritative tail (full-text mode);
+ * content, which we append at the end and reset the cursor.
- *   - otherwise → treat as an incremental delta and append.
+ *
+ * @param accumulated current merged text
+ * @param chunk       next chunk from the stream
+ * @param msg         the message being streamed (holds the replay cursor)
+ * @param isReasoning whether this targets the reasoning buffer or the content
+ * @returns the updated accumulated text
  */
-export function reconcileStream(accumulated: string, chunk: string): string {
+export function reconcileStream(
+  accumulated: string,
+  chunk: string,
+  msg: ChatMessageDisplay,
+  isReasoning: boolean,
+): string {
   if (chunk.length === 0) return accumulated;
-  if (accumulated.length === 0) return chunk;
+  const posRef = { pos: isReasoning ? msg._posReasoning : msg._posContent };
-  if (accumulated.length >= chunk.length && accumulated.startsWith(chunk)) {
+  if (accumulated.length === 0) {
-    // Shorter (or equal) replay of what we already hold — ignore.
+    accumulated = chunk;
+    posRef.pos = chunk.length;
+    writePos();
     return accumulated;
   }
-  if (chunk.length >= accumulated.length && chunk.startsWith(accumulated)) {
+  // 1) A chunk that continues the current replay cursor is a replayed segment
-    // Full-text mode: the chunk is the whole transcript so far (matching our
+  //    (it already exists earlier in the text) → skip it, advance the cursor.
-    // accumulated text plus more). Adopt it wholesale.
+  if (
-    return chunk;
+    posRef.pos > 0 &&
+    posRef.pos < accumulated.length &&
+    chunk.length <= accumulated.length - posRef.pos &&
+    accumulated.startsWith(chunk, posRef.pos)
+  ) {
+    posRef.pos += chunk.length;
+    writePos();
+    return accumulated;
+  }
+  // 2) A chunk that matches at the very start begins a NEW replay wave → skip
+  //    it, set the cursor to just past the matched prefix.
+  if (accumulated.startsWith(chunk)) {
+    posRef.pos = chunk.length;
+    writePos();
+    return accumulated;
+  }
+  // 3) Otherwise this is brand-new content: append it and reset the cursor so
+  //    the next wave starts fresh.
+  accumulated += chunk;
+  posRef.pos = -1;
+  writePos();
+  return accumulated;
+
+  function writePos(): void {
+    if (isReasoning) msg._posReasoning = posRef.pos;
+    else msg._posContent = posRef.pos;
   }
-  // Incremental delta mode: a new suffix. Append.
-  return accumulated + chunk;
 }
 
 /* ── AppStateRest ─────────────────────────────────────────────────── */
@@ -436,6 +567,8 @@ export interface AppStateRest {
   turnAbort: AbortController | null;
   /** Token usage from the most recent turn (surfaced in the status bar). */
   lastUsage: { tokensIn: number; tokensOut: number } | null;
+  /** Monotonic boot time (ms) used for the status-bar uptime + spinner. */
+  bootedAt: number;
 }
 
 export const DEFAULT_HELP_TEXT = ` Zesdex TUI — Keyboard Shortcuts
@@ -446,6 +579,7 @@ export const DEFAULT_HELP_TEXT = ` Zesdex TUI — Keyboard Shortcuts
   Ctrl+Y        Copy last assistant message
   Esc           Abort turn / Close overlay
   Tab           Autocomplete
+  Alt+D         Show git diff
 
  ─── Navigation ───
   ↑ / ↓         History browse / Overlay navigate
@@ -455,20 +589,32 @@ export const DEFAULT_HELP_TEXT = ` Zesdex TUI — Keyboard Shortcuts
 
  ─── Overlays ───
   /help         Show this help
-  /settings     Open settings overlay
+  /model        Pick provider → model (interactive)
+  /config       Settings menu (internet, review, provider)
   /todo         Open tasks (todo) overlay
   /usage        Open usage statistics
-  /bash         Open bash jobs overlay
+  /diff         Open git diff overlay
+  /plan         Open plan overlay
   /mcp          Open MCP server management
-  /model        Open model selector
   /compact      Compact conversation
   /clear        Clear transcript
-  /rewind       Rewind conversation history
+
+ ─── Menus (interactive pickers) ───
+  ↑/↓ · j/k     Move selection
+  PgUp/PgDn     Page     Home/End — jump
+  g / G         Top / Bottom
+  type a-z      Type-to-filter
+  Enter         Open / apply
+  Esc           Back / close
 
  ─── Editor Mode ───
   /edit <path>  Open file for inline editing
-  Ctrl+S        Save changes
+  Esc           Dismiss editor
-  Esc           Dismiss editor`;
+
+ ─── Legend ───
+  ● running    turn is being executed
+  ● idle       waiting for input
+  in:/out:     token usage of the latest turn`;
 
 /** Construct fresh TUI state. */
 export function createTuiState(
@@ -508,6 +654,7 @@ export function createTuiState(
     turnRunner: null,
     turnAbort: null,
     lastUsage: null,
+    bootedAt: Date.now(),
   };
 }
 

src/interfaces/tui/tui.test.ts

@@ -4,10 +4,11 @@
  */
 import { describe, test, expect } from "bun:test";
 import { parseCommand, applyCommand } from "./command.ts";
-import { createTuiState, pushTranscript, makeChatMessage, reconcileStream } from "./state.ts";
+import { createTuiState, pushTranscript, makeChatMessage, reconcileStream, type ChatMessageDisplay } from "./state.ts";
 import { applyAction } from "./action.ts";
 import { handleKey, decodeKey } from "./controller.ts";
 import { toControllerKey } from "./ui.tsx";
+import { modelProviderMenu, modelSubMenu, resolveMenu } from "./menu.ts";
 import { Roles } from "@zesdex/domain";
 
 function freshState() {
@@ -94,6 +95,38 @@ describe("handleKey", () => {
     expect(actions).toEqual([{ tag: "SubmitInput", text: "a" }]);
   });
 
+  test("autocomplete: partial /mo + Enter completes and runs /model", () => {
+    const s = freshState();
+    for (const c of "/mo") handleKey({ code: "Char", value: c, ctrl: false, alt: false }, s);
+    expect(s.input.autocompleteVisible).toBe(true);
+    expect(s.input.autocompleteCandidates).toContain("/model");
+    // A single Enter on a matching autocomplete completes the token AND runs
+    // the command (no unknown-command error, no clear, no double-Enter).
+    const first = handleKey({ code: "Enter", ctrl: false, alt: false }, s);
+    expect(s.input.buffer).toBe("/model");
+    expect(s.input.autocompleteVisible).toBe(false);
+    expect(first).toEqual([{ tag: "OpenMenuById", menu: "model" }]);
+  });
+
+  test("autocomplete: full /model hides dropdown so one Enter runs it", () => {
+    const s = freshState();
+    for (const c of "/model") handleKey({ code: "Char", value: c, ctrl: false, alt: false }, s);
+    expect(s.input.autocompleteVisible).toBe(false); // exact match → no dropdown
+    const actions = handleKey({ code: "Enter", ctrl: false, alt: false }, s);
+    expect(actions).toEqual([{ tag: "OpenMenuById", menu: "model" }]);
+  });
+
+  test("autocomplete: /edit path keeps its args on Enter (no clobber)", () => {
+    const s = freshState();
+    for (const c of "/edit src/a.ts") handleKey({ code: "Char", value: c, ctrl: false, alt: false }, s);
+    expect(s.input.autocompleteVisible).toBe(false); // /edit matches exactly too
+    const actions = handleKey({ code: "Enter", ctrl: false, alt: false }, s);
+    // Should dispatch the /edit command with path, preserving the typed args.
+    const opened = actions.map((a) => a.tag);
+    expect(opened).toContain("OpenEditor");
+    expect(s.input.buffer).toBe("/edit src/a.ts");
+  });
+
   test("Ctrl+Y copies last assistant message", () => {
     const s = freshState();
     pushTranscript(s, makeChatMessage(Roles.User, "hi"));
@@ -112,6 +145,84 @@ describe("handleKey", () => {
   });
 });
 
+describe("interactive menu stack (/model, /config)", () => {
+  test("/model command opens a menu (OpenMenuById → model)", () => {
+    const s = freshState();
+    s.appConfig = { providers: { claude: {}, zen: {}, router: {} } };
+    const actions = applyCommand(parseCommand("/model"));
+    expect(actions).toEqual([{ tag: "OpenMenuById", menu: "model" }]);
+    for (const a of actions) applyAction(s, a);
+    const frame = s.misc.menuStack[s.misc.menuStack.length - 1];
+    expect(frame?.id).toBe("model:provider");
+    expect(frame?.items).toEqual(["claude", "zen", "router"]);
+    expect(s.misc.overlay).toBe("menu");
+  });
+
+  test("/config command opens the settings menu", () => {
+    const s = freshState();
+    s.settings = { provider: "zen", internet_mode: "Off", review_enabled: true };
+    const actions = applyCommand(parseCommand("/config"));
+    applyAction(s, actions[0]!);
+    const frame = s.misc.menuStack[s.misc.menuStack.length - 1];
+    expect(frame?.id).toBe("config:root");
+    expect(frame?.items).toContain("Internet access");
+    expect(frame?.items).toContain("Provider / model");
+  });
+
+  test("arrow/jk navigation moves selection; Home/End/extended wraps", () => {
+    const s = freshState();
+    s.appConfig = { providers: { claude: {}, zen: {}, router: {} } };
+    applyAction(s, { tag: "OpenMenuById", menu: "model" });
+    applyAction(s, { tag: "MenuDown" });
+    expect(s.misc.menuStack.at(-1)?.selected).toBe(1);
+    applyAction(s, { tag: "MenuEnd" });
+    expect(s.misc.menuStack.at(-1)?.selected).toBe(2);
+    applyAction(s, { tag: "MenuDown" }); // wraps 2 → 0
+    expect(s.misc.menuStack.at(-1)?.selected).toBe(0);
+    applyAction(s, { tag: "MenuHome" });
+    expect(s.misc.menuStack.at(-1)?.selected).toBe(0);
+    const claudeIndex = 0;
+    void claudeIndex;
+  });
+
+  test("provider → model nested flow applies and pops the stack", () => {
+    const s = freshState();
+    s.appConfig = { providers: { claude: {}, zen: {}, router: {} } };
+    s.settings = { provider: "zen", model: "x" };
+    applyAction(s, { tag: "OpenMenuById", menu: "model" });
+    // The active provider (zen) is pre-selected when the menu opens.
+    const stack = () => s.misc.menuStack;
+    expect(stack().at(-1)?.items[0]).toBe("claude");
+    expect(stack().at(-1)?.selected).toBe(1); // zen
+    // Reuse a tiny helper that mirrors ui.tsx: apply every action from handleKey.
+    const drive = (k: Parameters<typeof handleKey>[0]) => {
+      for (const a of handleKey(k, s)) applyAction(s, a);
+    };
+    // Enter on the active provider → opens its model sub-menu.
+    drive({ code: "Enter", ctrl: false, alt: false });
+    expect(stack().length).toBe(2);
+    expect(stack().at(-1)?.id).toBe("model:zen");
+    expect(stack().at(-1)?.items.length).toBeGreaterThan(0);
+    // Enter on a model row applies zen/model and pops the whole stack.
+    drive({ code: "Enter", ctrl: false, alt: false });
+    expect(s.settings.provider).toBe("zen");
+    expect(s.misc.overlay as unknown as string).toBe("none");
+  });
+
+  test("type-to-filter narrows selection to the first match", () => {
+    const s = freshState();
+    s.appConfig = { providers: { claude: {}, zen: {}, router: {} } };
+    applyAction(s, { tag: "OpenMenuById", menu: "model" });
+    handleKey({ code: "Char", value: "r", ctrl: false, alt: false }, s);
+    const frame = s.misc.menuStack.at(-1)!;
+    expect(frame.filter).toBe("r");
+    expect(frame.selected).toBe(2); // "router" is the r match
+    // Backspace clears the filter.
+    handleKey({ code: "Backspace", ctrl: false, alt: false }, s);
+    expect(s.misc.menuStack.at(-1)?.filter).toBe("");
+  });
+});
+
 describe("decodeKey", () => {
   test("Enter decodes", () => {
     expect(decodeKey("\r")).toEqual({ code: "Enter", ctrl: false, alt: false });
@@ -163,32 +274,137 @@ describe("toControllerKey (OpenTUI adapter)", () => {
   });
 });
 
-describe("reconcileStream (full-text vs incremental chunk handling)", () => {
+describe("reconcileStream (9router sliding-window vs incremental)", () => {
-  test("appends incremental deltas", () => {
+  function msg(): ChatMessageDisplay {
-    let s = reconcileStream("", "Hello");
+    return makeChatMessage(Roles.Assistant, "");
-    s = reconcileStream(s, "!");
+  }
-    s = reconcileStream(s, " How");
+  function rec(
+    s: string,
+    chunk: string,
+    m: ChatMessageDisplay,
+  ): string {
+    return reconcileStream(s, chunk, m, false);
+  }
+
+  test("incremental deltas append in order", () => {
+    const m = msg();
+    let s = rec("", "Hello", m);
+    s = rec(s, "!", m);
+    s = rec(s, " How", m);
     expect(s).toBe("Hello! How");
   });
-  test("full-text mode: adopts a growing authoritative chunk", () => {
+
-    let s = reconcileStream("", "Hello");
+  test("real 9router wave: replayed segments never double", () => {
-    // 9router/proxy resends the *whole* transcript each chunk — adopt it.
+    // Real captured stream (story turn), each fresh token restarts a replay
-    s = reconcileStream(s, "Hello world");
+    // wave from the head: "Here's a 3-sent" "ence story" " about" " a robo" …
-    s = reconcileStream(s, "Hello world!");
+    const m = msg();
-    expect(s).toBe("Hello world!");
+    const segs: string[] = [
-  });
+      "Here's a 3-sent",
-  test("shorter/equal replay of the tail is ignored (no duplication)", () => {
+      "ence story",
-    const tail = "Hello world";
+      " about",
-    expect(reconcileStream(tail, tail)).toBe(tail);
+      " a robo",
-    expect(reconcileStream(tail, "Hello")).toBe(tail);
+      "t:\n\nA small repai",
-  });
+      "r robo",
-  test("partial-then-full garbled stream resolves to the full text", () => {
+      "t name",
-    let s = reconcileStream("", "Hello! 更新");
+      "d Bolt",
-    s = reconcileStream(s, "Hello! 更新 Keep going");
+    ];
-    expect(s).toBe("Hello! 更新 Keep going");
+    // Wave 1 (first two segments).
+    let s = "";
+    s = rec(s, segs[0]!, m);
+    s = rec(s, segs[1]!, m);
+    expect(s).toBe("Here's a 3-sentence story");
+    // Wave 2 replays seg0..seg1 then new seg2.
+    for (const g of segs.slice(0, 3)) s = rec(s, g, m);
+    expect(s).toBe("Here's a 3-sentence story about");
+    // Wave 3 replays seg0..seg2 then new seg3.
+    for (const g of segs.slice(0, 4)) s = rec(s, g, m);
+    expect(s).toBe("Here's a 3-sentence story about a robo");
+    // Wave 4 replays 0..3 then new seg4.
+    for (const g of segs.slice(0, 5)) s = rec(s, g, m);
+    expect(s).toBe("Here's a 3-sentence story about a robot:\n\nA small repai");
+    // Wave 5 replays 0..4 then new seg5.
+    for (const g of segs.slice(0, 6)) s = rec(s, g, m);
+    expect(s).toBe(
+      "Here's a 3-sentence story about a robot:\n\nA small repair robo",
+    );
+    // Wave 6 replays 0..5 then new seg6.
+    for (const g of segs.slice(0, 7)) s = rec(s, g, m);
+    // Final wave replays everything, no new content.
+    for (const g of segs.slice(0, 8)) s = rec(s, g, m);
+    expect(s).toBe(
+      "Here's a 3-sentence story about a robot:\n\nA small repair robot named Bolt",
+    );
   });
+
   test("empty chunk is a no-op", () => {
-    expect(reconcileStream("abc", "")).toBe("abc");
+    const m = msg();
+    expect(rec("abc", "", m)).toBe("abc");
+  });
+
+  test("1-char shared letter is NOT treated as a replay", () => {
+    // "Hi! " ends with a space; "How…" starts with 'H' — must append fully.
+    const m = msg();
+    let s = rec("", "Hi! ", m);
+    s = rec(s, "How can I help you today?", m);
+    expect(s).toBe("Hi! How can I help you today?");
+  });
+
+  test("'say hi' captured sequence resolves cleanly", () => {
+    const m = msg();
+    let s = "";
+    // Real captured: "", "", "Hi! 👋 ", "", "", "Hi! 👋 ", "How can I help you today?"
+    s = rec(s, "Hi! 👋 ", m);
+    s = rec(s, "Hi! 👋 ", m); // replay of head
+    s = rec(s, "How can I help you today?", m);
+    expect(s).toBe("Hi! 👋 How can I help you today?");
   });
 });
 
+
+describe("menu model lists (live vs fallback)", () => {
+  test("provider menu lists configured providers", () => {
+    const s = freshState();
+    s.appConfig = { providers: { zen: {}, router: {}, claude: {} } };
+    s.settings = { provider: "router" };
+    const frame = modelProviderMenu(s);
+    expect(frame.id).toBe("model:provider");
+    expect(frame.items).toEqual(["zen", "router", "claude"].map(() => expect.anything()).map((_, i) => ["zen","router","claude"][i]!));
+    expect(frame.selected).toBe(1); // router active
+  });
+
+  test("model sub-menu uses live /models list when available", () => {
+    const s = freshState();
+    s.misc.availableModels = ["gpt-5", "claude-sonnet-5", "deepseek-v3"];
+    s.settings = { provider: "zen", model: "claude-sonnet-5" };
+    const frame = modelSubMenu(s, "zen")!;
+    expect(frame.items).toEqual(["gpt-5", "claude-sonnet-5", "deepseek-v3"]);
+    expect(frame.selected).toBe(1); // current model highlighted
+    expect(frame.title).toContain("(3)");
+  });
+
+  test("model sub-menu falls back to a static catalog without a live list", () => {
+    const s = freshState();
+    s.misc.availableModels = [];
+    s.settings = { provider: "zen", model: "" };
+    const frame = modelSubMenu(s, "zen")!;
+    expect(frame.items.length).toBeGreaterThan(0);
+    expect(frame.hint).toContain("fallback");
+  });
+
+  test("resolving a live model row applies it and pops the stack", () => {
+    const s = freshState();
+    s.misc.availableModels = ["gpt-5", "claude-sonnet-5"];
+    s.appConfig = { providers: { zen: {} } };
+    s.settings = { provider: "zen", model: "claude-sonnet-5" };
+    applyAction(s, { tag: "OpenMenu", frame: modelProviderMenu(s) });
+    applyAction(s, { tag: "OpenMenu", frame: modelSubMenu(s, "zen")! });
+    // The sub-menu pre-selects the CURRENT model (claude-sonnet-5, index 1).
+    expect(s.misc.menuStack.at(-1)?.selected).toBe(1);
+    // Move to index 0 (gpt-5), then resolve Enter for the model row.
+    applyAction(s, { tag: "MenuHome" });
+    const acts = resolveMenu(s, "model:zen");
+    expect(acts).toEqual([{ tag: "MenuBackAll" }]);
+    expect(s.settings.provider).toBe("zen");
+    expect(s.settings.model).toBe("gpt-5");
+  });
+});

src/interfaces/tui/ui.tsx

@@ -5,10 +5,11 @@
  * input through the existing `handleKey` → `applyAction` pipeline.
  *
  * Layout (column flex over the full terminal):
- *   1. Header bar   — app · model · session        (blue bar)
+ *   1. Header bar    — logo · model · session · live turn spinner
- *   2. Transcript   — role-coloured chat scrollbox (flex-grow)
+ *   2. Transcript    — role-coloured chat scrollbox (markdown-lite, flex-grow)
- *   3. Input row    — `>` prompt + buffer          (same line)
+ *   3. Activity bar  — collapsed breadcrumb of the running turn's tools/agents
- *   4. Status bar   — turn state · token usage · hints
+ *   4. Input row     — `❯` prompt + buffer + cursor
+ *   5. Status bar    — turn state · token usage · uptime · toast · hints
  *
  * The logic layer (`state.ts`, `action.ts`, `command.ts`, `controller.ts`)
  * stays logic-pure and unit-tested; this file only reads it for display.
@@ -23,7 +24,7 @@ import {
 import type { KeyEvent as OpenTuiKey, CliRenderer } from "@opentui/core";
 import type { ReactNode } from "react";
 
-import type { AppStateRest } from "./state.ts";
+import type { AppStateRest, ChatMessageDisplay } from "./state.ts";
 import { pushTranscript, makeChatMessage, appendToLastTranscript } from "./state.ts";
 import { Roles } from "@zesdex/domain";
 import { handleKey, KeyCode } from "./controller.ts";
@@ -71,6 +72,10 @@ export function toControllerKey(key: MinimalKey): ZelKeyEvent {
     case "pagedown":
     case "pgdn":
       return { code: KeyCode.PageDown, ctrl, alt };
+    case "home":
+      return { code: KeyCode.Home, ctrl, alt };
+    case "end":
+      return { code: KeyCode.End, ctrl, alt };
     case "escape":
     case "esc":
       return { code: KeyCode.Esc, ctrl, alt };
@@ -84,13 +89,24 @@ export function toControllerKey(key: MinimalKey): ZelKeyEvent {
   }
 }
 
-/** Role label mapped to a colour for the transcript pane. */
+/* ── Palette ──────────────────────────────────────────────────────── */
-const roleColor: Record<string, string> = {
+
-  [Roles.User]: "cyan",
+/** Zesdex brand colours — a calm, readable "deep ocean" scheme. */
-  [Roles.Assistant]: "green",
+const C = {
-  [Roles.System]: "yellow",
+  bg: "#0d1117", // deep off-navy transcript ground
-  [Roles.Tool]: "magenta",
+  surface: "#161b22", // panel / secondary background
-};
+  border: "#2a3340", // dim hairline separators
+  fg: "#c9d1d9", // primary foreground
+  dim: "#6e7681", // muted text
+  cyan: "#58a6ff", // user / accents
+  green: "#3fb950", // assistant
+  yellow: "#d29922", // system
+  magenta: "#bc8cff", // agent / tools
+  red: "#f85149", // errors
+  headerA: "#0d47a1", // header gradient start (deep blue)
+  headerB: "#1f3a5f", // header gradient end
+  accent: "#58a6ff", // prompt / highlight
+} as const;
 
 /** Short uppercase role tag shown before each message line. */
 const roleTag: Record<string, string> = {
@@ -100,8 +116,13 @@ const roleTag: Record<string, string> = {
   [Roles.Tool]: "TOOL",
 };
 
-const oneLine = (s: string): string =>
+/** Distinct colour per role — used for badges + body tint. */
-  s.replace(/\s+$/g, "").replace(/\n+/g, "\n").trim();
+const roleColor: Record<string, string> = {
+  [Roles.User]: C.cyan,
+  [Roles.Assistant]: C.green,
+  [Roles.System]: C.yellow,
+  [Roles.Tool]: C.magenta,
+};
 
 /** Column-align a pair of header cells across the full width. */
 function spreadCenter(left: string, right: string, width: number): string {
@@ -109,6 +130,201 @@ function spreadCenter(left: string, right: string, width: number): string {
   return left + " ".repeat(gap) + right;
 }
 
+/** Pad a string to at least `n` chars, clipping if longer. */
+function pad(s: string, n: number): string {
+  return s.length >= n ? s : s + " ".repeat(n - s.length);
+}
+
+/* ── Spinner ──────────────────────────────────────────────────────── */
+
+const SPINNER_FRAMES = ["⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"];
+
+/** Pick a spinner glyph that animates with the tick counter. */
+function spinner(state: AppStateRest): string {
+  const frame = Math.floor(state.misc.tickCount / 2) % SPINNER_FRAMES.length;
+  return SPINNER_FRAMES[frame] ?? "⠋";
+}
+
+/* ── Markdown-lite rendering ──────────────────────────────────────── */
+
+/** One styled run of text within a wrapped line. */
+interface Run {
+  text: string;
+  fg?: string;
+  bold?: boolean;
+}
+
+/**
+ * Parse a modest Markdown subset into styled runs, then word-wrap them to a
+ * fixed width with a left indent. Keeps layout fully under our control (vs the
+ * self-drawing `<markdown>` node) so streaming + indent stay stable.
+ *
+ * Supported: **bold**, `inline code`, #headers, - bullets, > quotes, tables
+ * are untouched (rare in agent output). Code fences are rendered as
+ * single-colour blocks.
+ */
+function renderMarkdown(text: string, indent: number, width: number): Run[][] {
+  const rows: Run[][] = [];
+  const bodyWidth = Math.max(10, width - indent);
+  const wrap = new Wrap(bodyWidth);
+
+  for (const raw of stripFences(text).split("\n")) {
+    const line = raw.replace(/\s+$/, "");
+    if (line.trim().length === 0) {
+      wrap.break();
+      continue;
+    }
+    if (/^#{1,6}\s+/.test(line)) {
+      const seg: Run = { text: line.replace(/^#{1,6}\s+/, "").trim(), bold: true, fg: C.cyan };
+      wrap.add([seg]);
+      continue;
+    }
+    if (/^[-*]\s+/.test(line)) {
+      // Bullet: prefix glyph, then inline-parse the body so **bold**/`code`
+      // inside list items render too.
+      wrap.add([{ text: "• ", fg: C.fg }, ...inlineRuns(line.replace(/^[-*]\s+/, ""))]);
+      continue;
+    }
+    if (/^>\s?/.test(line)) {
+      wrap.add([{ text: "▍ ", fg: C.dim }, ...inlineRuns(line.replace(/^>\s?/, ""))]);
+      continue;
+    }
+    const runs = inlineRuns(line);
+    wrap.add(runs);
+  }
+  // Flatten wrapped rows into the output.
+  for (const row of wrap.rows()) {
+    rows.push(row);
+  }
+  return rows;
+}
+
+/** Render plain text (user / system / tool bodies) with wrapping + indent. */
+function renderPlain(text: string, indent: number, width: number): Run[][] {
+  const bodyWidth = Math.max(10, width - indent);
+  const wrap = new Wrap(bodyWidth);
+  for (const raw of text.split("\n")) {
+    wrap.add([{ text: raw.replace(/\s+$/, "") }]);
+  }
+  return wrap.rows();
+}
+
+/** A line-oriented wrapper that accumulates runs and breaks on width. */
+class Wrap {
+  private out: Run[][] = [];
+  private cur: Run[] = [];
+  private curLen = 0;
+  constructor(private width: number) {}
+  add(runs: Run[]): void {
+    for (const r of runs) this.pushRun(r);
+    this.break();
+  }
+  private pushRun(r: Run): void {
+    if (this.curLen + r.text.length > this.width && this.cur.length > 0) {
+      this.break();
+    }
+    if (r.text.length === 0) return;
+    this.cur.push(r);
+    this.curLen += r.text.length;
+  }
+  break(): void {
+    if (this.cur.length > 0) {
+      this.out.push(this.cur);
+      this.cur = [];
+      this.curLen = 0;
+    }
+  }
+  rows(): Run[][] {
+    this.break();
+    return this.out;
+  }
+}
+
+/** Tokenise a line into bold / inline-code / plain runs. */
+function inlineRuns(line: string): Run[] {
+  const runs: Run[] = [];
+  const re = /(\*\*[^*]+\*\*|`[^`]+`)/g;
+  let last = 0;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(line)) !== null) {
+    if (m.index > last) runs.push({ text: line.slice(last, m.index) });
+    const tok = m[0];
+    if (tok.startsWith("**")) {
+      runs.push({ text: tok.slice(2, -2), bold: true, fg: C.fg });
+    } else {
+      runs.push({ text: tok.slice(1, -1), fg: C.magenta });
+    }
+    last = m.index + tok.length;
+  }
+  if (last < line.length) runs.push({ text: line.slice(last) });
+  return runs.length ? runs : [{ text: line }];
+}
+
+/** Neutralise fenced code blocks so inline parsing stays clean (kept as plain). */
+function stripFences(text: string): string {
+  return text.replace(/```[a-zA-Z0-9]*\n?/g, "");
+}
+
+/* ── Transcript rendering ─────────────────────────────────────────── */
+
+/** One chat message rendered in the transcript. */
+function TranscriptMessage(props: {
+  msg: ChatMessageDisplay;
+  width: number;
+}): ReactNode {
+  const { msg, width } = props;
+  const color = roleColor[msg.role] ?? C.fg;
+  const tag = roleTag[msg.role] ?? msg.role?.toUpperCase() ?? "?";
+  const isAssistant = msg.role === Roles.Assistant;
+  // Assistant messages legitimately have no body while a tool call is being
+  // made (or mid-stream) — don't fill that with a misleading placeholder.
+  const content =
+    isAssistant && !msg.content.trim() ? "" : msg.content || "(empty)";
+
+  const rows: Run[][] =
+    msg.role === Roles.Assistant && content.length > 0
+      ? renderMarkdown(content, 5, width)
+      : renderPlain(content, 5, width);
+
+  return (
+    <box flexDirection="column" width="100%">
+      {/* Role badge header line */}
+      <text fg={color}>
+        <b>{`  ${pad(`[${tag}]`, 6)}`}</b>
+      </text>
+      {/* Body — each wrapped line is a <text> whose styled runs are <span>/<b> */}
+      {rows.map((row, i) => (
+        <text key={i} fg={C.fg}>
+          {row.map((r, j) => {
+            const text = r.text;
+            if (r.bold) {
+              return (
+                <b key={j} fg={r.fg ?? color}>
+                  {text}
+                </b>
+              );
+            }
+            return (
+              <span key={j} fg={r.fg ?? color}>
+                {text}
+              </span>
+            );
+          })}
+        </text>
+      ))}
+      {/* Live reasoning (dimmed, kept to one line) */}
+      {msg.reasoning.length > 0 && (
+        <text fg={C.dim}>{`    ⋯ ${oneLine(msg.reasoning)}`}</text>
+      )}
+      {/* Thin separator between messages */}
+      <text fg={C.border}>{`  ${"─".repeat(Math.max(0, Math.min(60, width - 4)))}`}</text>
+    </box>
+  );
+}
+
+const oneLine = (s: string): string =>
+  s.replace(/\s+$/g, "").replace(/\n+/g, "\n").trim();
+
 /**
  * Root OpenTUI component for the Zesdex TUI.
  * @param state       The mutable AppStateRest (mutated by applyAction).
@@ -121,17 +337,33 @@ export function ZesdexApp(props: { state: AppStateRest; onQuit: () => void }): R
   useKeyboard((key: OpenTuiKey) => {
     if (key.eventType === "release") return;
     const zk = toControllerKey(key);
+    let consumedByModal = false;
     try {
       const actions: Action[] = handleKey(zk, state);
       for (const act of actions) {
         if (act.tag === "QuitConfirm") {
           state.quit = true;
           onQuit();
+          consumedByModal = true;
           continue;
         }
+        // A slash-command that opens a modal (overlay / menu) "consumes" the
+        // input line — clear the buffer so `/model` doesn't linger. Autocomplete
+        // selection and normal submit must NOT be cleared (they own the buffer).
+        const opensModal =
+          act.tag === "OpenOverlay" ||
+          act.tag === "OpenMenuById" ||
+          act.tag === "OpenEditor" ||
+          act.tag === "ShowDiff" ||
+          act.tag === "ModelList";
+        if (opensModal && !state.input.autocompleteVisible) consumedByModal = true;
         applyAction(state, act);
       }
     } finally {
+      // Clear the input line only when a modal command consumed it this press.
+      if (consumedByModal) {
+        state.input.clear();
+      }
       // Force React to re-read the mutated `state` on the next frame.
       setTick((t: number) => t + 1);
     }
@@ -140,6 +372,7 @@ export function ZesdexApp(props: { state: AppStateRest; onQuit: () => void }): R
   const dimensions = useTerminalDimensionsHook();
   const width = dimensions.width ?? 80;
   const height = dimensions.height ?? 24;
+  const running = state.turnInFlightFlag.value;
 
   // Turn driver: when a turn is submitted, run it against the real agent and
   // stream every event into the transcript as it happens.
@@ -178,8 +411,10 @@ export function ZesdexApp(props: { state: AppStateRest; onQuit: () => void }): R
         state.turnInFlightFlag.value = false;
         state.turnAbort = null;
         state.abortFlag.value = false;
-        // Re-prime the next poll so queued turns run.
+        // Keep the last activity breadcrumb visible briefly, then clear it on
+        // the next turn start (see the start of poll()).
         setTimeout(() => {
+          state.misc.activity = [];
           turnRunning = false;
           setTick((x: number) => x + 1);
         }, 0);
@@ -189,191 +424,502 @@ export function ZesdexApp(props: { state: AppStateRest; onQuit: () => void }): R
     return () => clearInterval(timer);
   }, [state, setTick]);
 
-  const model = String(state.settings.model ?? state.settings.provider ?? "?");
+  // Tick the spinner + uptime while idle too (so ● running animates and the
-  const sessionId = String(state.settings.sessionId ?? state.sessionId ?? "");
+  // uptime clock in the status bar advances even between key presses).
-  const headerLeft = ` zesdex  ·  ${model} `;
+  useEffect(() => {
-  const headerRight = sessionId ? ` session:${sessionId} ` : "  ";
+    if (!running) return;
-  const header = spreadCenter(headerLeft, headerRight, width);
+    const id = setInterval(() => {
+      state.misc.tickCount += 1;
+      state.dirty = true;
+      setTick((t: number) => t + 1);
+    }, 120);
+    return () => clearInterval(id);
+  }, [running, state, setTick]);
 
-  const messages = state.transcriptCache.messages;
+  const overlayActive = state.misc.overlay !== "none";
-  const scrollFromEnd = messages.length - state.scroll.offset;
+  const menuActive = state.misc.menuStack.length > 0;
-  const visible = messages.slice(Math.max(0, scrollFromEnd - 40), scrollFromEnd);
 
-  const running = state.turnInFlightFlag.value;
+  // Modal overlays and interactive menus take the full body (below the header,
+  // above the status bar), replacing the transcript + input.
+  const body = menuActive ? (
+    <MenuView state={state} height={height} />
+  ) : overlayActive ? (
+    <OverlayView state={state} width={width} height={height} />
+  ) : (
+    <box flexDirection="column" flexGrow={1} flexShrink={1} minHeight={0}>
+      <scrollbox
+        flexGrow={1}
+        flexShrink={1}
+        flexBasis={0}
+        minHeight={0}
+        style={{ backgroundColor: C.bg }}
+        stickyScroll
+        stickyStart="bottom"
+      >
+        <TranscriptView state={state} width={width} />
+      </scrollbox>
+      {/* Dropdown (compact, sized to content) + activity + input, in order. The
+          input row is flexShrink:0 so the transcript scrollbox absorbs overflow
+          and the text area is never hidden. */}
+      {state.input.autocompleteVisible && <AutocompleteView state={state} />}
+      <ActivityBar state={state} running={running} />
+      <InputView state={state} width={width} />
+    </box>
+  );
 
   return (
     <box flexDirection="column" width={width} height={height}>
       {/* Header */}
-      <box backgroundColor="blue">
+      <HeaderView state={state} running={running} width={width} />
-        <text fg="white">{header}</text>
+      {body}
-      </box>
-
-      {/* Transcript pane */}
-      <scrollbox flexGrow={1} style={{ backgroundColor: "#0b0f14" }}>
-        {visible.map((m, i) => (
-          <TranscriptMessage key={`${m.timestamp}-${i}`} msg={m} width={width} />
-        ))}
-        {messages.length === 0 && (
-          <text fg="gray">  — no messages yet. Type a message below or /help. —</text>
-        )}
-      </scrollbox>
-
-      {/* Autocomplete dropdown (shown above the input row) */}
-      {state.input.autocompleteVisible && <AutocompleteView state={state} />}
-
-      {/* Overlay or input row */}
-      {state.misc.overlay !== "none" ? (
-        <OverlayView state={state} />
-      ) : (
-        <InputView state={state} width={width} />
-      )}
-
       {/* Status bar */}
-      <StatusBar state={state} running={running} />
+      <StatusBar state={state} running={running} width={width} />
     </box>
   );
 }
 
-/** Word-wrap a string to a fixed width (used for plain-text messages). */
+/* ── Header ───────────────────────────────────────────────────────── */
-function wrapText(text: string, width: number): string {
-  const lines: string[] = [];
-  for (const raw of text.split("\n")) {
-    // Never wrap an empty (or whitespace-only) line.
-    if (raw.length === 0) {
-      lines.push("");
-      continue;
-    }
-    let line = "";
-    for (const word of raw.split(/(\s+)/)) {
-      if (word === "\n") {
-        if (line) lines.push(line);
-        line = "";
-        continue;
-      }
-      if (line.length + word.length > width) {
-        if (line) lines.push(line);
-        line = word; // long word spills; next iteration handles it
-      } else {
-        line += word;
-      }
-    }
-    if (line) lines.push(line);
-  }
-  return lines.join("\n");
-}
 
-/** One chat message rendered in the transcript. */
+function HeaderView(props: {
-function TranscriptMessage(props: {
+  state: AppStateRest;
-  msg: AppStateRest["transcriptCache"]["messages"][number];
+  running: boolean;
   width: number;
 }): ReactNode {
-  const { msg, width } = props;
+  const { state, running, width } = props;
-  const color = roleColor[msg.role] ?? "white";
+  const model = String(state.settings.model ?? state.settings.provider ?? "?");
-  const tag = roleTag[msg.role] ?? msg.role?.toUpperCase() ?? "?";
+  const sessionId = String(state.settings.sessionId ?? state.sessionId ?? "");
-  // Assistant messages legitimately have no body while a tool call is being
+  const shortId = sessionId.length > 12 ? sessionId.slice(0, 11) : sessionId;
-  // made (or mid-stream) — don't fill that with a misleading placeholder.
+  const glyph = running ? spinner(state) : "◆";
-  const content =
+  const left = ` ${glyph} zesdex  ·  ${model} `;
-    msg.role === Roles.Assistant && !msg.content.trim()
+  const right = sessionId ? ` session:${shortId} ` : "  ";
-      ? ""
+  const header = spreadCenter(left, right, width);
-      : msg.content || "(empty)";
+  const bg = running ? C.headerA : C.headerB;
-
-  // Each message is a labelled block: a coloured role tag line followed by the
-  // word-wrapped body. Plain `<text>` is used (not `<markdown>`) so live
-  // streaming re-renders a single clean string — no overlap artifacts, and the
-  // body fills the full transcript width.
   return (
-    <box flexDirection="column" width="100%">
+    <box backgroundColor={bg}>
-      <text fg={color}>{` ${tag.padEnd(4)} `}</text>
+      <text fg={running ? C.green : "white"}>
-      {content.length > 0 && (
+        <b>{header}</b>
-        <text fg={color}>{wrapText(content, Math.max(20, width - 4))}</text>
-      )}
-      {msg.reasoning.length > 0 && (
-        <text fg="gray">{`  ⋯ ${oneLine(msg.reasoning)}`}</text>
-      )}
-    </box>
-  );
-}
-
-/** Bottom input row: `>` prompt + buffer on one line. */
-function InputView(props: { state: AppStateRest; width: number }): ReactNode {
-  const { state, width } = props;
-  const shown = state.input.buffer.slice(
-    Math.max(0, state.input.cursor - Math.max(10, width - 12)),
-    Math.max(0, state.input.cursor) + 40,
-  );
-  return (
-    <box flexDirection="row" width="100%">
-      <text fg="cyan">{"> "}</text>
-      <text fg="white">{shown}</text>
-    </box>
-  );
-}
-
-/** Autocomplete / file-mention dropdown rendered as a bordered panel. */
-function AutocompleteView(props: { state: AppStateRest }): ReactNode {
-  const { state } = props;
-  const candidates = state.input.autocompleteCandidates;
-  const kind = state.input.autocompleteKind;
-  const title = kind === "command" ? "commands" : "files";
-  return (
-    <box width="100%" style={{ border: true, padding: 0 }}>
-      <text fg="magenta">
-        {` ${title}  (↑/↓ navigate, Tab select, Esc close)\n`}
       </text>
-      {candidates.slice(0, 10).map((c, i) => (
+    </box>
-        <text key={c} fg={i === state.input.autocompleteIdx ? "cyan" : "white"}>
+  );
-          {i === state.input.autocompleteIdx ? "▸ " : "  "}
+}
-          {c}
+
-          {i < candidates.slice(0, 10).length - 1 ? "\n" : ""}
+/* ── Transcript ───────────────────────────────────────────────────── */
-        </text>
+
+function TranscriptView(props: { state: AppStateRest; width: number }): ReactNode {
+  const { state, width } = props;
+  const messages = state.transcriptCache.messages;
+  const scrollFromEnd = messages.length - state.scroll.offset;
+  const visible = messages.slice(
+    Math.max(0, scrollFromEnd - 40),
+    scrollFromEnd,
+  );
+
+  if (messages.length === 0) {
+    return (
+      <text fg={C.dim}>
+        {`\n  ─  no messages yet. Type a message below  (or /help)  ─`}
+      </text>
+    );
+  }
+  return (
+    <box flexDirection="column">
+      {visible.map((m, i) => (
+        <TranscriptMessage key={`${m.timestamp}-${i}`} msg={m} width={width} />
       ))}
     </box>
   );
 }
 
-/** Full-screen modal overlays (help, diff, editor, model selector, etc.). */
+/* ── Activity bar (running turn) ──────────────────────────────────── */
-function OverlayView(props: { state: AppStateRest }): ReactNode {
+
-  const { state } = props;
+function ActivityBar(props: { state: AppStateRest; running: boolean }): ReactNode {
-  const name = state.misc.overlay;
+  const { state, running } = props;
-  let body = "";
+  const activity = state.misc.activity;
-  if (name === "quit_confirm") body = "Quit? (y=yes, n=no)  Ctrl+C again / /quit to confirm";
+  if (!running || activity.length === 0) return null;
-  else if (name === "help") body = state.helpText;
+  const line = ` ${spinner(state)} ${activity.map((a) => label(a)).join("  ·  ")} `;
-  else if (name === "diff")
-    body = state.misc.diffContent
-      .split("\n")
-      .slice(state.misc.diffScroll, state.misc.diffScroll + 20)
-      .join("\n");
-  else body = `Overlay: ${name}  (Esc to close)`;
   return (
-    <box flexDirection="column" width="100%" style={{ border: true, padding: 1 }}>
+    <box flexDirection="column" width="100%" style={{ backgroundColor: C.surface }}>
-      <text fg="yellow">{body}</text>
+      <text fg={C.magenta}>
-      <text fg="gray">
+        {line}
-        {name === "quit_confirm" ? "y=yes  n=no  Esc=cancel" : "Esc to close"}
+      </text>
+    </box>
+  );
+
+  function label(a: { label: string; kind: string; detail?: string; count?: number }): string {
+    const sym = a.kind === "error" ? "✕" : a.kind === "agent" ? "⧉" : "⚙";
+    const n = a.count && a.count > 1 ? `×${a.count}` : "";
+    return `${sym} ${a.label}${n}`;
+  }
+}
+
+/* ── Input row ────────────────────────────────────────────────────── */
+
+function InputView(props: { state: AppStateRest; width: number }): ReactNode {
+  const { state, width } = props;
+  const buffer = state.input.buffer;
+  const cursor = state.input.cursor;
+  const shown = buffer.slice(
+    Math.max(0, cursor - Math.max(10, width - 14)),
+    cursor + 40,
+  );
+  const hasContent = buffer.length > 0;
+  return (
+    <box flexDirection="column" width="100%" flexShrink={0}>
+      <text fg={C.border}>{`─`.repeat(Math.max(0, width))}</text>
+      <box flexDirection="row" width="100%">
+        <text fg={C.accent}>
+          <b>{`  ❯ `}</b>
+        </text>
+        {hasContent ? (
+          <text fg={C.fg}>{shown}</text>
+        ) : (
+          <text fg={C.dim}>Ask Zesdex to help…</text>
+        )}
+        <text fg={C.accent}>{` ▏`}</text>
+      </box>
+    </box>
+  );
+}
+
+/* ── Autocomplete ─────────────────────────────────────────────────── */
+
+function AutocompleteView(props: { state: AppStateRest }): ReactNode {
+  const { state } = props;
+  const candidates = state.input.autocompleteCandidates;
+  const kind = state.input.autocompleteKind;
+  const title = kind === "command" ? "commands" : "files";
+  const selected = Math.min(state.input.autocompleteIdx, candidates.length - 1);
+  const visible = candidates.slice(0, 10);
+  // Render the whole panel as ONE multiline <text> — this fixes the OpenTUI
+  // quirk where sibling <text> children overlap. The box is sized to its
+  // content (flexShrink 0, explicit height) so it never squashes the input row.
+  const lines = [
+    ` ${title}`,
+    ...visible.map((c, i) => (i === selected ? `▸ ${c}` : `  ${c}`)),
+    `  ↑/↓ navigate · Tab accept · Esc close`,
+  ];
+  const body = lines.join("\n");
+  return (
+    <box
+      width="100%"
+      flexDirection="column"
+      flexGrow={0}
+      flexShrink={0}
+      style={{ border: true, borderColor: C.border, backgroundColor: C.surface, padding: 1 }}
+    >
+      <text fg={kind === "command" ? C.accent : C.magenta} height={lines.length}>
+        {body}
       </text>
     </box>
   );
 }
 
-/** Bottom status bar: turn state · token usage · live toast/hints. */
+/* ── Overlays ─────────────────────────────────────────────────────── */
-function StatusBar(props: { state: AppStateRest; running: boolean }): ReactNode {
+
-  const { state, running } = props;
+function OverlayView(props: {
+  state: AppStateRest;
+  width: number;
+  height: number;
+}): ReactNode {
+  const { state } = props;
+  const name = state.misc.overlay;
+
+  switch (name) {
+    case "quit_confirm":
+      return (
+        <ConfirmOverlay
+          title="Quit Zesdex?"
+          body="  Are you sure you want to quit?\n\n  Any in-flight turn will be aborted."
+          hint="y=yes  n=no  Esc=cancel"
+        />
+      );
+    case "clear_confirm":
+      return (
+        <ConfirmOverlay
+          title="Clear transcript?"
+          body="  This clears the visible conversation from the transcript view."
+          hint="y=yes  n=no  Esc=cancel"
+        />
+      );
+    case "help":
+      return <HelpOverlay state={state} />;
+    case "diff":
+      return <DiffOverlay state={state} />;
+    case "model_selector":
+      return <ModelSelectorOverlay state={state} />;
+    case "editor":
+      return <EditorOverlay state={state} />;
+    default:
+      return <PlaceholderOverlay name={name} state={state} />;
+  }
+}
+
+function ConfirmOverlay(props: { title: string; body: string; hint: string }): ReactNode {
+  const { title, body, hint } = props;
+  return (
+    <OverlayFrame title={title}>
+      <text fg={C.fg}>{body}</text>
+      <text fg={C.dim}>{`  ${hint}`}</text>
+    </OverlayFrame>
+  );
+}
+
+function HelpOverlay(props: { state: AppStateRest }): ReactNode {
+  const { state } = props;
+  return (
+    <OverlayFrame title="Keyboard Shortcuts">
+      <text fg={C.fg}>{state.helpText}</text>
+      <text fg={C.dim}>{"  Esc or /help to close"}</text>
+    </OverlayFrame>
+  );
+}
+
+function DiffOverlay(props: { state: AppStateRest }): ReactNode {
+  const { state } = props;
+  const lines = state.misc.diffContent.split("\n");
+  const start = state.misc.diffScroll;
+  const slice = lines.slice(start, start + 18);
+  return (
+    <OverlayFrame title="Git Diff">
+      <scrollbox flexGrow={1} stickyScroll stickyStart="bottom">
+        {slice.map((l, i) => {
+          const fg = l.startsWith("+")
+            ? C.green
+            : l.startsWith("-")
+              ? C.red
+              : l.startsWith("@@")
+                ? C.cyan
+                : C.dim;
+          return (
+            <text key={i} fg={fg}>
+              {l.length ? l : " "}
+            </text>
+          );
+        })}
+      </scrollbox>
+      <text fg={C.dim}>
+        {`  ↑/↓ / PgUp/PgDn scroll · Esc close  — ${Math.min(start + 1, lines.length)}/${lines.length}`}
+      </text>
+    </OverlayFrame>
+  );
+}
+
+function ModelSelectorOverlay(props: { state: AppStateRest }): ReactNode {
+  const { state } = props;
+  const names = Object.keys(state.appConfig.providers);
+  const list = names.length ? names : ["claude", "zen", "router"];
+  const idx = Math.min(Math.max(0, state.misc.selectedIndex), list.length - 1);
+  const current = String(state.settings.provider ?? "");
+  const selected = idx;
+  const body = list
+    .map((p, i) => {
+      const isSel = i === selected;
+      const active = p === current;
+      const suffix = active ? "  (active)" : "";
+      return `${isSel ? "▸ " : "  "}${pad(p, 15)}${suffix}`;
+    })
+    .join("\n");
+  return (
+    <OverlayFrame title="Select Model / Provider">
+      <text fg={C.fg}>
+        {body}
+      </text>
+      <text fg={C.dim}>{"  ↑/↓ select · Enter apply · Esc close"}</text>
+    </OverlayFrame>
+  );
+}
+
+function EditorOverlay(props: { state: AppStateRest }): ReactNode {
+  const { state } = props;
+  const ed = state.misc.editor;
+  if (!ed) return <OverlayFrame title="Editor"><text fg={C.fg}>No buffer.</text></OverlayFrame>;
+  const lines = ed.content.split("\n");
+  const start = Math.max(0, ed.cursor - 5);
+  const slice = lines.slice(start, start + 12);
+  return (
+    <OverlayFrame title={`Edit ${ed.path}`}>
+      {slice.map((l, i) => (
+        <text key={i} fg={C.fg}>
+          {`${String(start + i + 1).padStart(3)} | ${l}`}
+        </text>
+      ))}
+      <text fg={C.dim}>
+        {"  type to edit · Enter newline · Esc save & close"}
+      </text>
+    </OverlayFrame>
+  );
+}
+
+function PlaceholderOverlay(props: { name: string; state: AppStateRest }): ReactNode {
+  const { name, state } = props;
+  const content =
+    name === "todo"
+      ? state.misc.todoContent
+      : name === "plan"
+        ? state.misc.planContent
+        : "";
+  const summary =
+    name === "todo" ? "todo list" :
+    name === "plan" ? "plan outline" :
+    name === "usage" ? "usage statistics" :
+    name === "mcp" ? "MCP server management" :
+    name === "settings" ? "settings" :
+    name === "bash" ? "bash jobs" :
+    name === "rewind" ? "rewind history" :
+    name === "learning" ? "learned patterns" :
+    "overlay";
+  return (
+    <OverlayFrame title={`${name} — ${summary}`}>
+      {content ? (
+        <text fg={C.fg}>{content}</text>
+      ) : (
+        <text fg={C.dim}>{`  ${summary}: nothing yet.`}</text>
+      )}
+      <text fg={C.dim}>{"  Esc to close"}</text>
+    </OverlayFrame>
+  );
+}
+
+function OverlayFrame(props: { title: string; children: ReactNode }): ReactNode {
+  const { title, children } = props;
+  return (
+    <box
+      width="100%"
+      flexDirection="column"
+      flexGrow={1}
+      flexShrink={0}
+      style={{
+        border: true,
+        borderColor: C.accent,
+        padding: 1,
+        backgroundColor: C.surface,
+      }}
+      title={title}
+      titleAlignment="center"
+    >
+      {children}
+    </box>
+  );
+}
+
+/* ── Interactive menu ─────────────────────────────────────────────── */
+
+/** Modal picker for the menu stack — like Claude Code / opencode selectors. */
+function MenuView(props: { state: AppStateRest; height: number }): ReactNode {
+  const { state, height } = props;
+  const stack = state.misc.menuStack;
+  const frame = stack[stack.length - 1];
+  if (!frame) return null;
+  const filter = (frame.filter ?? "").toLowerCase();
+  // Rows that match the live filter (empty filter = all rows).
+  const rawMatched = frame.items
+    .map((it, i) => ({ it, i }))
+    .filter((r) => !filter || r.it.toLowerCase().includes(filter));
+  // Fit the visible rows to the modal height (top rows take priority so the
+  // selected row stays anchored near the top).
+  const maxRows = Math.max(4, height - 8);
+  const visibleStart = Math.max(0, Math.min(rawMatched.length - maxRows, frame.selected));
+  const matched = rawMatched.slice(visibleStart, visibleStart + maxRows);
+  const detailOf = frame.details;
+
+  // Breadcrumb: titles from the stack, e.g. "Model › pick a model (zen)".
+  const crumb = stack.map((f) => f.title).join(" › ");
+
+  return (
+    <OverlayFrame title={crumb}>
+      {/* Quick filter bar at the top */}
+      <text fg={C.dim}>{filter ? `  filter: ${frame.filter}   ` : "  type to filter…"}</text>
+      {matched.map(({ it, i }) => {
+        const isSel = i === frame.selected;
+        const detail = detailOf?.[i] ?? "";
+        return (
+          <box key={i} flexDirection="row" width="100%">
+            <text fg={isSel ? C.accent : C.fg} flexGrow={1}>
+              {isSel ? <b>{`${"▸ "}${it}`}</b> : `  ${it}`}
+            </text>
+            {detail && (
+              <text fg={C.dim} flexShrink={0}>
+                {detail}
+              </text>
+            )}
+          </box>
+        );
+      })}
+      {matched.length === 0 && (
+        <text fg={C.red}>  no matches — type less or press Backspace.</text>
+      )}
+      <text fg={C.dim}>
+        {`  ${frame.hint ?? ""}`}
+      </text>
+      {frame.items.length > 10 && (
+        <text fg={C.dim}>
+          {`  showing ${matched.length}/${frame.items.length} · PgUp/PgDn scroll`}
+        </text>
+      )}
+    </OverlayFrame>
+  );
+}
+
+/* ── Status bar ───────────────────────────────────────────────────── */
+
+function StatusBar(props: {
+  state: AppStateRest;
+  running: boolean;
+  width: number;
+}): ReactNode {
+  const { state, running, width } = props;
   const left = running ? "● running" : "● idle";
   const usage = state.lastUsage;
   const usageStr = usage
     ? `in:${usage.tokensIn} · out:${usage.tokensOut}`
     : "in:– · out:–";
+
+  // Uptime since boot.
+  const uptime = fmtUptime((Date.now() - state.bootedAt) / 1000);
+
   // Autocomplete hint for the active input.
   const ac =
     state.input.autocompleteVisible &&
     state.input.autocompleteCandidates[state.input.autocompleteIdx] != null
       ? ` [tab: ${state.input.autocompleteCandidates[state.input.autocompleteIdx]}]`
       : "";
+
   // Most recent live toast (expires via MiscState).
   const toasts = state.misc.toasts;
   const toast = toasts.length > 0 ? ` (${toasts[toasts.length - 1]?.message})` : "";
+
   const hints = running ? "Esc abort" : "/help · Ctrl+C quit · ↑ history";
-  const cell = ` ${left}   ${usageStr}${ac}${toast}   ${hints} `;
+
-  return <text fg={toast ? "yellow" : running ? "green" : "gray"}>{cell}</text>;
+  const statusColor = toast
+    ? C.yellow
+    : running
+      ? C.green
+      : C.dim;
+
+  const activity = running ? ` ${spinner(state)}` : "";
+
+  const cell = ` ${left}${activity}   ${usageStr}   ${uptime}${ac}${toast}   ${hints} `;
+  return (
+    <box backgroundColor={C.surface}>
+      <text fg={statusColor}>{trimTo(cell, width)}</text>
+    </box>
+  );
+}
+
+/** Clip a status line to the available width (avoid overflow). */
+function trimTo(s: string, width: number): string {
+  if (s.length <= width) return s;
+  return s.slice(0, Math.max(0, width - 1)) + "…";
+}
+
+/** Format seconds as a compact human uptime (e.g. 1m23s). */
+function fmtUptime(sec: number): string {
+  const s = Math.floor(sec);
+  const m = Math.floor(s / 60);
+  const h = Math.floor(m / 60);
+  if (h > 0) return `${h}h${String(m % 60).padStart(2, "0")}m`;
+  if (m > 0) return `${m}m${String(s % 60).padStart(2, "0")}s`;
+  return `${s}s`;
 }
 
 /** Boot the OpenTUI renderer and mount the Zesdex app. */
@@ -400,3 +946,4 @@ export async function bootTui(state: AppStateRest): Promise<void> {
     }, 60);
   });
 }
+