Skip to content

A self-improving RLM agent for coding workflows and long-running autonomous tasks.

Rust

Keep the session thinking level for /btw side questions (#2234)

Sebastian Müller5255a6b88002
diff --git a/packages/coding-agent/.changes/btw-side-question-thinking.md b/packages/coding-agent/.changes/btw-side-question-thinking.md
new file mode 100644
index 0000000000..a418c747b0
--- /dev/null
+++ b/packages/coding-agent/.changes/btw-side-question-thinking.md
@@ -0,0 +1 @@
+- Fixed `/btw` side questions discarding the main conversation's prompt cache by lowering the reasoning level: side questions now keep the session's thinking level, so providers whose cache keys include thinking parameters reuse the cached conversation.
diff --git a/packages/coding-agent/src/core/side-question.ts b/packages/coding-agent/src/core/side-question.ts
index 53cac1ac9f..0b7ab10353 100644
--- a/packages/coding-agent/src/core/side-question.ts
+++ b/packages/coding-agent/src/core/side-question.ts
@@ -6,7 +6,6 @@ import {
type ProviderRetryPolicy,
} from "./provider-retry.js";
import { unwrapSemanticEdgeStreamFn } from "./semantic-edges.js";
-import { getAuxiliaryThinkingLevel } from "./thinking-levels.js";
export type SideQuestionStatus = "running" | "complete" | "cancelled" | "error";
@@ -97,7 +96,9 @@ export function startSideQuestion(
model,
systemPrompt: parent.state.systemPrompt,
messages: [...structuredClone(parent.state.messages), ...previousTurnMessages],
- thinkingLevel: getAuxiliaryThinkingLevel(model, parent.state.thinkingLevel),
+ // Anthropic message-level caching keys on the thinking parameters, so a
+ // different level here would re-read the whole cloned conversation.
+ thinkingLevel: parent.state.thinkingLevel,
serviceTier: parent.state.serviceTier,
// Providers serialize tool declarations ahead of the cached prefix, so an
// empty list would miss the main cache; execution is blocked in beforeToolCall.
diff --git a/packages/coding-agent/test/suite/regressions/6010-auxiliary-reasoning.test.ts b/packages/coding-agent/test/suite/regressions/6010-auxiliary-reasoning.test.ts
index a6231804cd..9f0b3fad19 100644
--- a/packages/coding-agent/test/suite/regressions/6010-auxiliary-reasoning.test.ts
+++ b/packages/coding-agent/test/suite/regressions/6010-auxiliary-reasoning.test.ts
@@ -70,11 +70,13 @@ describe("auxiliary reasoning settings", () => {
parent.state.thinkingLevel = testCase.parentLevel;
const parentMessages = structuredClone(parent.state.messages);
const observed: SimpleStreamOptions[] = [];
+ // Side questions keep the session level (cache identity); the first two calls clamp.
+ const expectedLevels: ThinkingLevel[] = [testCase.expectedLevel, testCase.expectedLevel, testCase.parentLevel];
harness.setResponses(
- [JSON.stringify(proposal), JSON.stringify(review), "Side answer"].map((text) => (_context, options) => {
+ [JSON.stringify(proposal), JSON.stringify(review), "Side answer"].map((text, index) => (_context, options) => {
const request = options as SimpleStreamOptions;
observed.push(request);
- if (request.reasoning !== testCase.expectedLevel) {
+ if (request.reasoning !== expectedLevels[index]) {
return fauxAssistantMessage("", {
stopReason: "error",
errorMessage: `Unsupported auxiliary reasoning: ${request.reasoning}`,
@@ -131,7 +133,7 @@ describe("auxiliary reasoning settings", () => {
noRetry,
).done;
expect(events.at(-1)).toMatchObject({ status: "complete", answer: "Side answer" });
- expect(observed.map((options) => options.reasoning)).toEqual(Array(3).fill(testCase.expectedLevel));
+ expect(observed.map((options) => options.reasoning)).toEqual(expectedLevels);
expect(observed[0].maxTokens).toBe(testCase.expectedLevel === "off" ? 32_000 : model.maxTokens);
expect(observed[1].maxTokens).toBe(testCase.expectedLevel === "off" ? 4_096 : model.maxTokens);
expect(parent.state.thinkingLevel).toBe(testCase.parentLevel);