feat(tui): implement message compaction and summarization logic for efficient conversation handling

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

@@ -8,6 +8,9 @@ import {
   isToolFailure,
   deriveVerifyCommand,
   AgentTurnServiceImpl,
+  takeRecentTail,
+  reduceMessagesToDigest,
+  compactMessagesWithAi,
 } from "./turn_service.ts";
 import {
   type ChatMessage,
@@ -264,3 +267,118 @@ describe("runTurn self-review", () => {
     expect(events.some((e) => e.kind === "review_usage")).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,
@@ -40,6 +41,15 @@ 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"]);
 
@@ -250,8 +260,75 @@ async function executeToolCallsInParallel(
 /* -------------------------------------------------------------------------- */
 
 /**
- * Compact oversized conversation history using AI summarisation. At most once
- * per turn. Keeps the last COMPACT_KEEP_TAIL messages.
+ * Split history into a recent tail (kept verbatim, bound by CHAR budget so
+ * 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[],
@@ -259,17 +336,35 @@ export async function compactMessagesWithAi(
 ): Promise<void> {
   if (messages.length <= COMPACT_KEEP_TAIL + 2) return;
 
-  const splitIdx = messages.length - COMPACT_KEEP_TAIL;
-  const evicted = messages.splice(0, splitIdx);
+  const { tail, evicted } = takeRecentTail(messages);
+  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.";
-    messages.unshift(systemMessage(`[AI Summary of Previous Conversation]\n${summaryText.trim()}`));
+    const summaryText = message.content?.trim() ?? "";
+    const rebuilt: ChatMessage[] = [];
+    if (summaryText !== "") {
+      rebuilt.push(systemMessage(`[AI Summary of Previous Conversation]\n${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);
   }
 }