fix(compaction): do not forward agent tools to summarization

packages/coding-agent/src/core/extensions/builtin/compaction/changes.md

@@ -1,5 +1,16 @@
 # Builtin compaction extension changes
 
+## Summarization must not forward agent tools (2026-08-21)
+
+### What changed
+- `getSummarizationTools()` now returns `[]` via `compactionSummarizationTools()` instead of copying `pi.getActiveTools()`.
+- The summarizer remains a prose-only LLM call; `snapshot.tools` is empty so models cannot end the summary with a bare `toolUse`.
+
+### Why
+- Forwarding the live tool list made tool-preferring models (observed: openai-codex/gpt-5.6-sol) finish summarization with `stopReason: toolUse` and no text.
+- `getSummaryText()` only joins `content.type === "text"`, so that becomes `SummaryGenerationError("empty-summary")` and compaction is rejected while tokens stay above the threshold.
+- Observed on session `01a01fae` at 257k tokens: auto, retry, and manual `/compact` all rejected with empty-summary `toolUse`. The 2026-07-21 note already documented this failure mode.
+
 ## Skip Cursor compaction while the session is not idle (2026-08-19)
 
 Blocking and generated apply refuse `cursor` / `cursor-cli-oauth` when `!ctx.isIdle()`. Mid-run Cursor compact poisons `conversationId`. Idle `agent_end` / `pre_prompt` still compact.

packages/coding-agent/src/core/extensions/builtin/compaction/index.ts

@@ -13,6 +13,7 @@ import type {
 } from "../../types.ts";
 import * as checkpointState from "./checkpoint-state.ts";
 import * as breaker from "./circuit-breaker.ts";
+import { compactionSummarizationTools } from "./summarization-tools.ts";
 import {
 	BUILTIN_CONTEXT_REDUCTION_OPTIONS,
 	reduceContextMessages,
@@ -211,16 +212,7 @@ export default function compactionExtension(
 	const getLogger = (ctx: ExtensionContext): CompactionLogger => (logger ??= createCompactionLogger(ctx.agentDir));
 
 	function getSummarizationTools(): Tool[] {
-		if (typeof pi.getAllTools !== "function" || typeof pi.getActiveTools !== "function") return [];
-		try {
-			const definitionsByName = new Map(pi.getAllTools().map((tool) => [tool.name, tool]));
-			return pi.getActiveTools().flatMap((name) => {
-				const tool = definitionsByName.get(name);
-				return tool ? [{ name: tool.name, description: tool.description, parameters: tool.parameters }] : [];
-			});
-		} catch {
-			return [];
-		}
+		return compactionSummarizationTools();
 	}
 
 	let idleWarmupTimer: ReturnType<typeof setTimeout> | undefined;

packages/coding-agent/src/core/extensions/builtin/compaction/summarization-tools.ts

@@ -0,0 +1,11 @@
+import type { Tool } from "@earendil-works/pi-ai";
+
+/**
+ * Summarization is a prose-only LLM call. Forwarding the live agent tool list
+ * lets tool-preferring models (observed: openai-codex/gpt-5.6-sol) end the
+ * summary with `stopReason: toolUse` and no text, which senpi rejects as
+ * `empty-summary` and leaves the session above the compaction threshold.
+ */
+export function compactionSummarizationTools(_liveTools?: Tool[]): Tool[] {
+	return [];
+}

packages/coding-agent/test/compaction/summarization-no-tools.test.ts

@@ -0,0 +1,17 @@
+import { describe, expect, it } from "vitest";
+import type { Tool } from "@earendil-works/pi-ai";
+import { compactionSummarizationTools } from "../../src/core/extensions/builtin/compaction/summarization-tools.ts";
+
+describe("compactionSummarizationTools", () => {
+	it("returns no tools even when the live agent list is non-empty", () => {
+		const live = [
+			{
+				name: "bash",
+				description: "run a command",
+				parameters: { type: "object", properties: {} },
+			},
+		] as Tool[];
+		expect(compactionSummarizationTools(live)).toEqual([]);
+		expect(compactionSummarizationTools()).toEqual([]);
+	});
+});