0.4.5: constant system prompt per session, no more full re-reads every turn

CHANGELOG.md

@@ -3,6 +3,26 @@
 All notable changes to LocalCode will be documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
+## 0.4.5 — 2026-09-26
+
+### Fixed
+
+- **Every turn after the first is much faster.** llama-server reuses its cache
+  only for an identical prompt prefix, and the system prompt was changing
+  between steps: the workspace block was added after the first tool call of a
+  turn and dropped again on the next, the planning rule followed the same
+  switch, and the open-todo list was rewritten into the system prompt on every
+  update. Each change made the server re-read the whole conversation, twice
+  per turn (15-23 s at 30k tokens on an M5). The system prompt is now constant
+  for a session, and todos travel with the user turn. Measured on Qwen3.8 27B:
+  requests after the first tool call re-read 23-310 tokens (0.3-0.8 s)
+  instead of 6-7k (10-15 s).
+- Title generation and other short side requests no longer land on the
+  conversation's server slot when the server has several.
+- The code-intelligence panel said "Starts automatically for your project's
+  language", which was never true: localcode does not download language
+  servers on its own. It now says "Not running · /lsp to set up".
+
 ## 0.4.4 — 2026-09-26
 
 ### Fixed

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.4.4"
+version = "0.4.5"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

src/localcode/__init__.py

@@ -8,4 +8,4 @@
 try:
     __version__ = _pkg_version("localcode")
 except PackageNotFoundError:  # not installed (e.g. raw source checkout)
-    __version__ = "0.4.4"
+    __version__ = "0.4.5"

src/localcode/ui/FORK_COMMIT

@@ -1 +1 @@
-b88aa9395bb0cbc3aa9bf44a04e36ed0dea0f0bc
+f65ef33d195169f749bf406b50cdde8c819253f8

src/localcode/ui/plugin/localcode.ts

@@ -437,10 +437,8 @@ const LocalcodePlugin: Plugin = async ({ client, directory }) => {
     try {
       const r = await fetch(`${control}/status`, { signal: AbortSignal.timeout(1_000) });
       const j = (await r.json()) as { current?: string; state?: string };
-      loadedName = { name: j.state === "ready" && j.current ? j.current : loadedName.name, at: Date.now() };
-    } catch {
-      loadedName = { ...loadedName, at: Date.now() };
-    }
+      if (j.state === "ready" && j.current) loadedName = { name: j.current, at: Date.now() };
+    } catch {}
     return loadedName.name;
   }
 
@@ -509,10 +507,15 @@ const LocalcodePlugin: Plugin = async ({ client, directory }) => {
           if (output.system[i].includes("__pending__")) output.system[i] = output.system[i].replaceAll("__pending__", loaded);
         }
       }
-      if (!workspaceActive || (input.agent && input.agent !== "build")) return;
+      if (input.agent && input.agent !== "build") return;
+      // The system prompt must be IDENTICAL for every request of a session: it
+      // is the cached prefix. This rule used to be added only once
+      // workspaceActive flipped (after the first tool call of each turn), so the
+      // prompt changed mid-turn and the server re-read the whole conversation
+      // once per turn (15-23 s at 30k tokens). It is inert in plain chat (the
+      // gates check workspaceActive in code), so always add it. The open-todo
+      // list is not put here either: it rides on the user turn (chat.message).
       output.system.push(PLANNING_RULE);
-      const open = renderTodos(todos);
-      if (open) output.system.push(open);
     },
 
     "chat.message": async (input, output) => {
@@ -532,6 +535,10 @@ const LocalcodePlugin: Plugin = async ({ client, directory }) => {
         turnStartedAt = Date.now() - 1000;
         plateau = new PlateauTracker(directory); plateauStopped = false; seenSteps.clear();
       }
+      // Open todos travel with the turn (user message or nudge), never in the
+      // system prompt, so the cached prefix stays stable across todo updates.
+      const open = renderTodos(todos);
+      if (open) output.parts.push({ type: "text", text: open, synthetic: true } as any);
     },
 
     "tool.execute.after": async (input, output) => {

tests/ui/prompt-scope.test.ts

@@ -4,7 +4,11 @@ import { tmpdir } from "node:os";
 import { join } from "node:path";
 import Plugin from "../../src/localcode/ui/plugin/localcode";
 
-test("completion rules stay out of compaction and other helper prompts", async () => {
+// The system prompt is llama-server's cached prefix: it must be identical for
+// every request of a session, so the completion rules are present for the build
+// agent from the first request on (never toggled by workspace activity) and the
+// open-todo list travels on the user turn, not in the system prompt.
+test("completion rules stay out of helper prompts and are constant for the build agent", async () => {
   const directory = mkdtempSync(join(tmpdir(), "prompt-scope-"));
   try {
     const hooks = await (Plugin as any)({ directory, client: {} });
@@ -13,19 +17,17 @@ test("completion rules stay out of compaction and other helper prompts", async (
       await hooks["experimental.chat.system.transform"]({ agent }, output);
       expect(output.system).toEqual(["Summarize only."]);
     }
-    const initial = { system: [] as string[] };
-    await hooks["experimental.chat.system.transform"]({ agent: "build" }, initial);
-    expect(initial.system).toEqual([]);
-    await hooks["tool.execute.before"]({ tool: "websearch" }, { args: { query: "general research" } });
-    await hooks["experimental.chat.system.transform"]({ agent: "build" }, initial);
-    expect(initial.system).toEqual([]);
+    const first = { system: [] as string[] };
+    await hooks["experimental.chat.system.transform"]({ agent: "build" }, first);
+    expect(first.system.join("\n")).toContain("requested multi-step workspace changes");
     await hooks["tool.execute.before"]({ tool: "read" }, { args: { filePath: "." } });
-    const output = { system: [] as string[] };
-    await hooks["experimental.chat.system.transform"]({ agent: "build" }, output);
-    expect(output.system.join("\n")).toContain("requested multi-step workspace changes");
+    const afterTool = { system: [] as string[] };
+    await hooks["experimental.chat.system.transform"]({ agent: "build" }, afterTool);
+    expect(afterTool.system).toEqual(first.system);
     await hooks["chat.message"]({ sessionID: "test", agent: "build" }, { parts: [{ type: "text", text: "a different topic" }] });
-    const next = { system: [] as string[] };
-    await hooks["experimental.chat.system.transform"]({ agent: "build" }, next);
-    expect(next.system).toEqual([]);
+    const nextTurn = { system: [] as string[] };
+    await hooks["experimental.chat.system.transform"]({ agent: "build" }, nextTurn);
+    expect(nextTurn.system).toEqual(first.system);
+    expect(nextTurn.system.join("\n")).not.toContain("YOUR OPEN TODOS");
   } finally { rmSync(directory, { recursive: true, force: true }); }
 });