Promote v0.17.1 to main (compaction on 32k-window models)

packages/redrob/src/session/overflow.ts

@@ -95,7 +95,13 @@ export function usable(input: {
    * either way and the threshold applies to what is left.
    */
   const output = ProviderTransform.maxOutputTokens(input.model, input.outputTokenMax)
-  const headroom = Math.max(0, budget - Math.min(COMPACTION_BUFFER, output))
+  /*
+   * At most a quarter of the window. A 32,768-token model with the full 20,000 reserve was left
+   * 8,937 tokens, less than the system prompt and tool list alone, so every turn overflowed: a
+   * one-line question was summarised and auto-continued five times in one `run`.
+   */
+  const reserve = Math.min(COMPACTION_BUFFER, output, Math.floor(budget / 4))
+  const headroom = Math.max(0, budget - reserve)
   const windowLimit = Math.floor((headroom * clamped) / 100)
   /*
    * The cheaper of the two limits wins. `reserved` above returns before this on purpose: a caller who set

packages/redrob/test/session/threshold-and-cost.test.ts

@@ -37,8 +37,10 @@ describe("compaction threshold", () => {
 
   it("keeps room for the reply, whatever the threshold", () => {
     // A threshold of 100 must not mean "compact once the window is full", which is the failure the
-    // reserve existed to avoid.
-    expect(source).toContain("budget - Math.min(COMPACTION_BUFFER, output)")
+    // reserve existed to avoid. The reserve is capped at a quarter of the window so a 32k model is not
+    // left with less room than its own system prompt.
+    expect(source).toContain("Math.min(COMPACTION_BUFFER, output, Math.floor(budget / 4))")
+    expect(source).toContain("budget - reserve")
   })
 
   it("treats a published input cap of 0 as absent, not as a cap of zero", () => {