fix(opencode): fill the context meter during v2 turns (#5710)

OpenCode v2 sessions never reported context used or the model's context
limit, so the composer meter was an empty ring during a turn and vanished
when it ended. Each finished step now reports its context against the
selected model's limit.

Closes #5701

Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Matt Cowger <matt@cowger.us>

packages/server/src/server/agent/providers/opencode/v2/session.ts

@@ -35,6 +35,7 @@ import { features } from "./configuration.js";
 import { commands } from "./commands.js";
 import { messages } from "./history.js";
 import { SessionPermissions } from "./permissions.js";
+import { SessionUsage } from "./usage.js";
 
 export class OpenCodeV2Session implements AgentSession {
   readonly provider = "opencode";
@@ -45,6 +46,7 @@ export class OpenCodeV2Session implements AgentSession {
   private readonly permissions: SessionPermissions;
   private readonly turns: SessionTurns;
   private readonly children: SessionChildren;
+  private readonly usage: SessionUsage;
   private stream: Promise<void> | null = null;
   private streamAbort = new AbortController();
   private exited = false;
@@ -101,6 +103,17 @@ export class OpenCodeV2Session implements AgentSession {
           await this.permissions.respondToPermission(request.id, { behavior: "deny" });
       },
     });
+    this.usage = new SessionUsage({
+      client: () => this.client,
+      cwd: config.cwd,
+      info: () => this.info,
+      emit: (event) => this.emit(event),
+      reportError: (error) =>
+        this.logger.warn(
+          { error: toDiagnosticErrorMessage(error) },
+          "OpenCode context usage update failed",
+        ),
+    });
   }
   get id() {
     return this.info.id;
@@ -465,6 +478,7 @@ export class OpenCodeV2Session implements AgentSession {
       return;
     }
     this.permissions.observe(event);
+    this.usage.observe(event);
     this.turns.observe(event);
     this.scheduleReconcile();
   }

packages/server/src/server/agent/providers/opencode/v2/usage.test.ts

@@ -0,0 +1,100 @@
+import type { ModelInfo, TokenUsageInfo } from "@opencode/client";
+import { describe, expect, test } from "vitest";
+import { OpenCodeV2AgentClient } from "./agent.js";
+import { V2Harness } from "../test-utils/v2-harness.js";
+import { createTestLogger } from "../../../../../test-utils/test-logger.js";
+import type { AgentUsage } from "../../../agent-sdk-types.js";
+
+function model(context: number): ModelInfo {
+  return {
+    id: "model",
+    modelID: "model",
+    providerID: "test",
+    name: "Model",
+    capabilities: { tools: true, input: ["text"], output: ["text"] },
+    variants: [],
+    time: { released: 1 },
+    cost: [],
+    status: "active",
+    enabled: true,
+    limit: { context, output: 10_000 },
+  };
+}
+
+function stepEnded(harness: V2Harness, id: string, tokens: TokenUsageInfo) {
+  harness.push({
+    id,
+    created: 2,
+    type: "session.step.ended",
+    durable: { aggregateID: "session", seq: 1, version: 1 },
+    data: {
+      sessionID: "session",
+      assistantMessageID: "answer",
+      finish: "tool-calls",
+      cost: 0.1,
+      tokens,
+    },
+  });
+}
+
+describe("OpenCode v2 context usage", () => {
+  test("reports each step's context against the model limit, in step order", async () => {
+    const harness = new V2Harness();
+    harness.info.model = { providerID: "test", id: "model" };
+    harness.info.tokens = {
+      input: 900,
+      output: 90,
+      reasoning: 0,
+      cache: { read: 40, write: 0 },
+    };
+    harness.info.cost = 0.5;
+    // Each lookup answers faster than the one before it, so a lookup per step would
+    // deliver step 1's usage after step 2's.
+    const lookupDelays = [50, 0];
+    harness.api.model.list = async (input) => {
+      await new Promise((resolve) => setTimeout(resolve, lookupDelays.shift() ?? 0));
+      return {
+        location: { directory: input.location.directory },
+        data: [model(200_000)],
+      };
+    };
+    const client = new OpenCodeV2AgentClient({
+      logger: createTestLogger(),
+      runtime: harness.runtime,
+    });
+    const session = await client.createSession({
+      provider: "opencode",
+      cwd: "/tmp/project",
+    });
+    const usage: AgentUsage[] = [];
+    session.subscribe((event) => {
+      if (event.type === "usage_updated") usage.push(event.usage);
+    });
+    try {
+      stepEnded(harness, "step-1", {
+        input: 100,
+        output: 10,
+        reasoning: 5,
+        cache: { read: 1_000, write: 20 },
+      });
+      stepEnded(harness, "step-2", {
+        input: 300,
+        output: 10,
+        reasoning: 5,
+        cache: { read: 1_000, write: 20 },
+      });
+      await expect.poll(() => usage.length).toBe(2);
+      expect(usage.map((entry) => entry.contextWindowUsedTokens)).toEqual([1_135, 1_335]);
+      expect(usage[1]).toEqual({
+        inputTokens: 900,
+        outputTokens: 90,
+        cachedInputTokens: 40,
+        totalCostUsd: 0.5,
+        contextWindowUsedTokens: 1_335,
+        contextWindowMaxTokens: 200_000,
+      });
+    } finally {
+      await session.close();
+    }
+  });
+});

packages/server/src/server/agent/providers/opencode/v2/usage.ts

@@ -0,0 +1,61 @@
+import type { ModelRef, OpenCodeEvent, SessionInfo, TokenUsageInfo } from "@opencode/client";
+import type { AgentStreamEvent } from "../../../agent-sdk-types.js";
+import type { V2Api } from "./api.js";
+import { usageFromV2 } from "./mapping.js";
+
+interface UsageOptions {
+  client(): V2Api;
+  cwd: string;
+  info(): SessionInfo;
+  emit(event: AgentStreamEvent): void;
+  reportError(error: unknown): void;
+}
+
+// A step is one model call, so its tokens are the context that call used.
+function contextWindowUsedTokens(tokens: TokenUsageInfo) {
+  return tokens.input + tokens.output + tokens.reasoning + tokens.cache.read + tokens.cache.write;
+}
+
+export class SessionUsage {
+  private readonly limits = new Map<string, Promise<number | undefined>>();
+  private reports: Promise<void> = Promise.resolve();
+  constructor(private readonly options: UsageOptions) {}
+  observe(event: OpenCodeEvent) {
+    if (event.type !== "session.step.ended") return;
+    const { tokens } = event.data;
+    // Reports stay in step order while a model's context limit is still being looked up.
+    this.reports = this.reports
+      .then(() => this.report(tokens))
+      .catch((error: unknown) => this.options.reportError(error));
+  }
+  private async report(tokens: TokenUsageInfo) {
+    const model = this.options.info().model;
+    const contextWindowMaxTokens = model ? await this.contextLimit(model) : undefined;
+    this.options.emit({
+      type: "usage_updated",
+      provider: "opencode",
+      usage: {
+        ...usageFromV2(this.options.info()),
+        contextWindowUsedTokens: contextWindowUsedTokens(tokens),
+        ...(contextWindowMaxTokens === undefined ? {} : { contextWindowMaxTokens }),
+      },
+    });
+  }
+  private contextLimit(model: ModelRef) {
+    const key = `${model.providerID}/${model.id}`;
+    const cached = this.limits.get(key);
+    if (cached) return cached;
+    const limit = this.options
+      .client()
+      .model.list({ location: { directory: this.options.cwd } })
+      .then(
+        (catalog) =>
+          catalog.data.find(
+            (entry) => entry.providerID === model.providerID && entry.id === model.id,
+          )?.limit.context,
+      );
+    this.limits.set(key, limit);
+    limit.catch(() => this.limits.delete(key));
+    return limit;
+  }
+}