fix(glm): stabilize reasoning and context reporting (#3615)

Author: tatoalo

packages/agent/src/adapters/claude/claude-agent.streamed-text.test.ts

@@ -149,7 +149,17 @@ function textDelta(sessionId: string, text: string) {
   };
 }
 
-function assistantMessage(sessionId: string, apiId: string, text: string) {
+function assistantMessage(
+  sessionId: string,
+  apiId: string,
+  text: string,
+  usage?: {
+    input_tokens: number;
+    output_tokens: number;
+    cache_read_input_tokens: number;
+    cache_creation_input_tokens: number;
+  },
+) {
   return {
     type: "assistant",
     parent_tool_use_id: null,
@@ -159,6 +169,20 @@ function assistantMessage(sessionId: string, apiId: string, text: string) {
       id: apiId,
       role: "assistant",
       content: [{ type: "text", text }],
+      ...(usage ? { usage } : {}),
+    },
+  };
+}
+
+function compactBoundary(sessionId: string) {
+  return {
+    type: "system",
+    subtype: "compact_boundary",
+    session_id: sessionId,
+    uuid: "compact-1",
+    compact_metadata: {
+      trigger: "auto",
+      pre_tokens: 434_000,
     },
   };
 }
@@ -192,6 +216,23 @@ function messageChunkTexts(
     .map((update) => update?.content?.text ?? "");
 }
 
+function usageUpdates(
+  calls: ClientMocks["sessionUpdate"]["mock"]["calls"],
+): Array<{ used: number; size: number }> {
+  return calls.flatMap(([call]) => {
+    const update = (
+      call as {
+        update?: { sessionUpdate?: string; used?: number; size?: number };
+      }
+    ).update;
+    return update?.sessionUpdate === "usage_update" &&
+      typeof update.used === "number" &&
+      typeof update.size === "number"
+      ? [{ used: update.used, size: update.size }]
+      : [];
+  });
+}
+
 describe("ClaudeAcpAgent.prompt — streamed assistant text wiring", () => {
   beforeEach(() => {
     vi.clearAllMocks();
@@ -243,6 +284,48 @@ describe("ClaudeAcpAgent.prompt — streamed assistant text wiring", () => {
     ]);
   });
 
+  it("does not replace known context usage with incomplete gateway snapshots", async () => {
+    const { agent, client } = makeAgent();
+    const sessionId = "s-context-usage";
+    const { query, input } = installFakeSession(agent, sessionId);
+    const session = (
+      agent as unknown as {
+        session: { contextUsed?: number; lastContextWindowSize?: number };
+      }
+    ).session;
+    session.contextUsed = 434_000;
+    session.lastContextWindowSize = 1_000_000;
+    vi.mocked(query.getContextUsage).mockRejectedValue(
+      new Error("context usage unavailable"),
+    );
+
+    const promptPromise = agent.prompt({
+      sessionId,
+      prompt: [{ type: "text", text: "continue" }],
+    });
+    await tick();
+
+    await echoUserMessage(query, input);
+    await send(query, messageStart(sessionId, "msg-context"));
+    await send(query, compactBoundary(sessionId));
+    await send(
+      query,
+      assistantMessage(sessionId, "msg-context", "done", {
+        input_tokens: 440_000,
+        output_tokens: 100,
+        cache_read_input_tokens: 0,
+        cache_creation_input_tokens: 0,
+      }),
+    );
+    await send(query, resultSuccess(sessionId));
+    await promptPromise;
+
+    const updates = usageUpdates(client.sessionUpdate.mock.calls);
+    expect(updates.length).toBeGreaterThan(0);
+    expect(updates.every(({ used }) => used >= 434_000)).toBe(true);
+    expect(updates.every(({ size }) => size === 1_000_000)).toBe(true);
+  });
+
   it("keeps the original turn open until a pending steer is consumed", async () => {
     const { agent, client } = makeAgent();
     const sessionId = "s-steer-ordering";

packages/agent/src/adapters/claude/claude-agent.ts

@@ -651,6 +651,21 @@ export class ClaudeAcpAgent extends BaseAcpAgent {
       };
     };
 
+    const recordContextUsage = (nextTotal: number): boolean => {
+      if (nextTotal <= 0 || nextTotal === lastAssistantTotalUsage) {
+        return false;
+      }
+      const knownTotal = Math.max(
+        lastAssistantTotalUsage ?? 0,
+        session.contextUsed ?? 0,
+      );
+      if (nextTotal < knownTotal) {
+        return false;
+      }
+      lastAssistantTotalUsage = nextTotal;
+      return true;
+    };
+
     const resetTurnScratch = () => {
       lastAssistantTotalUsage = null;
       lastRefusalExplanation = null;
@@ -832,16 +847,19 @@ export class ClaudeAcpAgent extends BaseAcpAgent {
                 fetchContextUsedTokens(query, this.logger),
                 cancelController.signal,
               );
-              lastAssistantTotalUsage =
-                usedTokens.result === "success" ? (usedTokens.value ?? 0) : 0;
-              await this.client.sessionUpdate({
-                sessionId,
-                update: {
-                  sessionUpdate: "usage_update",
-                  used: lastAssistantTotalUsage,
-                  size: windowSize(),
-                },
-              });
+              if (usedTokens.result === "success" && usedTokens.value != null) {
+                lastAssistantTotalUsage = usedTokens.value;
+                session.contextUsed = usedTokens.value;
+                session.contextSize = windowSize();
+                await this.client.sessionUpdate({
+                  sessionId,
+                  update: {
+                    sessionUpdate: "usage_update",
+                    used: lastAssistantTotalUsage,
+                    size: windowSize(),
+                  },
+                });
+              }
             }
             if (message.subtype === "commands_changed") {
               session.knownSlashCommands = collectKnownSlashCommands(
@@ -1201,8 +1219,7 @@ export class ClaudeAcpAgent extends BaseAcpAgent {
                 lastStreamUsage.cache_read_input_tokens +
                 lastStreamUsage.cache_creation_input_tokens;
 
-              if (nextTotal !== lastAssistantTotalUsage) {
-                lastAssistantTotalUsage = nextTotal;
+              if (recordContextUsage(nextTotal)) {
                 await this.client.sessionUpdate({
                   sessionId,
                   update: {
@@ -1297,21 +1314,23 @@ export class ClaudeAcpAgent extends BaseAcpAgent {
                 cache_read_input_tokens: number | null;
                 cache_creation_input_tokens: number | null;
               };
-              lastAssistantTotalUsage =
+              const nextTotal =
                 (usage.input_tokens ?? 0) +
                 (usage.output_tokens ?? 0) +
                 (usage.cache_read_input_tokens ?? 0) +
                 (usage.cache_creation_input_tokens ?? 0);
 
-              await this.client.sessionUpdate({
-                sessionId,
-                update: {
-                  sessionUpdate: "usage_update",
-                  used: lastAssistantTotalUsage,
-                  size: windowSize(),
-                  cost: null,
-                },
-              });
+              if (recordContextUsage(nextTotal)) {
+                await this.client.sessionUpdate({
+                  sessionId,
+                  update: {
+                    sessionUpdate: "usage_update",
+                    used: nextTotal,
+                    size: windowSize(),
+                    cost: null,
+                  },
+                });
+              }
             }
 
             const result = await handleUserAssistantMessage(message, context);

packages/agent/src/adapters/claude/session/models.test.ts

@@ -95,6 +95,13 @@ describe("model capability flags", () => {
       xhighEffort: false,
       mcpInjection: false,
     },
+    {
+      modelId: "@cf/zai-org/glm-5.2",
+      oneMContext: false,
+      effort: true,
+      xhighEffort: false,
+      mcpInjection: true,
+    },
   ])(
     "$modelId capability flags",
     ({ modelId, oneMContext, effort, xhighEffort, mcpInjection }) => {
@@ -118,6 +125,7 @@ describe("resolveEffortForModel", () => {
     ["claude-opus-4-7", undefined, "high"],
     ["claude-sonnet-4-6", undefined, "high"],
     ["claude-sonnet-5", undefined, "high"],
+    ["@cf/zai-org/glm-5.2", undefined, "high"],
     // Models without effort support stay unset (SDK disables thinking).
     ["claude-haiku-4-5", undefined, undefined],
     ["claude-opus-4-6", undefined, undefined],
@@ -139,20 +147,12 @@ describe("getEffortOptions", () => {
     expect(getEffortOptions("claude-opus-4-6")).toBeNull();
   });
 
-  it("returns low/medium/high for effort-supporting models", () => {
-    const opts = getEffortOptions("claude-sonnet-4-6");
-    expect(opts?.map((o) => o.value)).toEqual(["low", "medium", "high"]);
-  });
-
-  it("appends xhigh and max for xhigh-supporting models", () => {
-    const opts = getEffortOptions("claude-opus-4-7");
-    expect(opts?.map((o) => o.value)).toEqual([
-      "low",
-      "medium",
-      "high",
-      "xhigh",
-      "max",
-    ]);
+  it.each([
+    ["claude-sonnet-4-6", ["low", "medium", "high"]],
+    ["claude-opus-4-7", ["low", "medium", "high", "xhigh", "max"]],
+    ["@cf/zai-org/glm-5.2", ["high", "max"]],
+  ])("returns the exact effort levels for %s", (modelId, expected) => {
+    expect(getEffortOptions(modelId)?.map((o) => o.value)).toEqual(expected);
   });
 });
 

packages/agent/src/adapters/claude/session/models.ts

@@ -33,23 +33,27 @@ export function supports1MContext(modelId: string): boolean {
   return MODELS_WITH_1M_CONTEXT.has(modelId);
 }
 
-const MODELS_WITH_EFFORT = new Set([
-  "claude-opus-4-7",
-  "claude-opus-4-8",
-  "claude-sonnet-4-6",
-  "claude-sonnet-5",
-  "claude-fable-5",
-]);
-
-const MODELS_WITH_XHIGH_EFFORT = new Set([
-  "claude-opus-4-7",
-  "claude-opus-4-8",
-  "claude-sonnet-5",
-  "claude-fable-5",
-]);
+const STANDARD_EFFORT_LEVELS: readonly EffortLevel[] = [
+  "low",
+  "medium",
+  "high",
+];
+const EXTENDED_EFFORT_LEVELS: readonly EffortLevel[] = [
+  ...STANDARD_EFFORT_LEVELS,
+  "xhigh",
+  "max",
+];
+const MODEL_EFFORT_LEVELS: Readonly<Record<string, readonly EffortLevel[]>> = {
+  "claude-opus-4-7": EXTENDED_EFFORT_LEVELS,
+  "claude-opus-4-8": EXTENDED_EFFORT_LEVELS,
+  "claude-sonnet-4-6": STANDARD_EFFORT_LEVELS,
+  "claude-sonnet-5": EXTENDED_EFFORT_LEVELS,
+  "claude-fable-5": EXTENDED_EFFORT_LEVELS,
+  "@cf/zai-org/glm-5.2": ["high", "max"],
+};
 
 export function supportsEffort(modelId: string): boolean {
-  return MODELS_WITH_EFFORT.has(modelId);
+  return MODEL_EFFORT_LEVELS[modelId] !== undefined;
 }
 
 export function resolveEffortForModel(
@@ -61,7 +65,7 @@ export function resolveEffortForModel(
 }
 
 export function supportsXhighEffort(modelId: string): boolean {
-  return MODELS_WITH_XHIGH_EFFORT.has(modelId);
+  return MODEL_EFFORT_LEVELS[modelId]?.includes("xhigh") ?? false;
 }
 
 const MODELS_TO_EXCLUDE_MCP_TOOLS = new Set(["claude-haiku-4-5"]);
@@ -82,27 +86,23 @@ export function fastModeStateEnabled(state: string | undefined): boolean {
 }
 
 interface EffortOption {
-  value: string;
+  value: EffortLevel;
   name: string;
 }
 
-export function getEffortOptions(modelId: string): EffortOption[] | null {
-  if (!supportsEffort(modelId)) return null;
-
-  const options: EffortOption[] = [
-    { value: "low", name: "Low" },
-    { value: "medium", name: "Medium" },
-    { value: "high", name: "High" },
-  ];
-
-  if (supportsXhighEffort(modelId)) {
-    options.push(
-      { value: "xhigh", name: "Extra High" },
-      { value: "max", name: "Max" },
-    );
-  }
+const EFFORT_LABELS: Record<EffortLevel, string> = {
+  low: "Low",
+  medium: "Medium",
+  high: "High",
+  xhigh: "Extra High",
+  max: "Max",
+};
 
-  return options;
+export function getEffortOptions(modelId: string): EffortOption[] | null {
+  const levels = MODEL_EFFORT_LEVELS[modelId];
+  return (
+    levels?.map((value) => ({ value, name: EFFORT_LABELS[value] })) ?? null
+  );
 }
 
 // Model alias resolution — lets callers use human-friendly aliases like

packages/agent/src/adapters/reasoning-effort.test.ts

@@ -39,4 +39,16 @@ describe("isSupportedReasoningEffort", () => {
       isSupportedReasoningEffort("claude", "claude-sonnet-4-6", "xhigh"),
     ).toBe(false);
   });
+
+  it.each([
+    ["high", true],
+    ["max", true],
+    ["low", false],
+    ["medium", false],
+    ["xhigh", false],
+  ])("validates GLM 5.2 effort %s", (effort, expected) => {
+    expect(
+      isSupportedReasoningEffort("claude", "@cf/zai-org/glm-5.2", effort),
+    ).toBe(expected);
+  });
 });