chore(main): release 0.30.0 (#573)

🤖 I have created a release *beep* *boop*
---


##
[0.30.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.29.2...v0.30.0)
(2026-04-20)


### Features

* Update to claude-agent-sdk 0.2.114
([#572](https://github.com/agentclientprotocol/claude-agent-acp/issues/572))
([e9dd452](e9dd45207d))

---
This PR was generated with [Release
Please](https://github.com/googleapis/release-please). See
[documentation](https://github.com/googleapis/release-please#release-please).

Co-authored-by: acp-release-bot[bot] <246668977+acp-release-bot[bot]@users.noreply.github.com>

.github/workflows/publish.yml

@@ -34,77 +34,6 @@ jobs:
       release_created: ${{ steps.release.outputs.release_created }}
       tag_name: ${{ steps.release.outputs.tag_name }}
 
-  build-binaries:
-    name: Build (${{ matrix.target }})
-    runs-on: ubuntu-latest
-    needs: [release-please]
-    # TODO: Re-enable once Bun SDK compatibility is fixed
-    if: false
-    permissions:
-      contents: read
-    timeout-minutes: 15
-
-    strategy:
-      fail-fast: true
-      matrix:
-        include:
-          - target: bun-linux-x64
-            artifact: claude-agent-acp
-            archive: claude-agent-acp-linux-x64.tar.gz
-          - target: bun-linux-arm64
-            artifact: claude-agent-acp
-            archive: claude-agent-acp-linux-arm64.tar.gz
-          - target: bun-linux-x64-musl
-            artifact: claude-agent-acp
-            archive: claude-agent-acp-linux-x64-musl.tar.gz
-          - target: bun-linux-arm64-musl
-            artifact: claude-agent-acp
-            archive: claude-agent-acp-linux-arm64-musl.tar.gz
-          - target: bun-darwin-x64
-            artifact: claude-agent-acp
-            archive: claude-agent-acp-darwin-x64.zip
-          - target: bun-darwin-arm64
-            artifact: claude-agent-acp
-            archive: claude-agent-acp-darwin-arm64.zip
-          - target: bun-windows-x64
-            artifact: claude-agent-acp.exe
-            archive: claude-agent-acp-windows-x64.zip
-          - target: bun-windows-arm64
-            artifact: claude-agent-acp.exe
-            archive: claude-agent-acp-windows-arm64.zip
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
-
-      - uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6
-        with:
-          bun-version: latest
-
-      - name: Install dependencies
-        run: bun install
-
-      - name: Build binary
-        run: |
-          bun build src/index.ts \
-            --compile \
-            --target ${{ matrix.target }} \
-            --define 'process.env.CLAUDE_AGENT_ACP_IS_SINGLE_FILE_BUN="true"' \
-            --outfile ${{ matrix.artifact }}
-
-      - name: Compress binary
-        run: |
-          if [[ "${{ matrix.archive }}" == *.zip ]]; then
-            zip ${{ matrix.archive }} ${{ matrix.artifact }}
-          else
-            tar czf ${{ matrix.archive }} ${{ matrix.artifact }}
-          fi
-
-      - name: Upload artifact
-        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a
-        with:
-          name: ${{ matrix.archive }}
-          path: ${{ matrix.archive }}
-
   publish-npm:
     name: Publish to npm
     runs-on: ubuntu-latest
@@ -126,31 +55,3 @@ jobs:
       - run: npm ci
       - run: npm run build
       - run: npm publish
-
-  release:
-    name: Create Release
-    runs-on: ubuntu-latest
-    needs: [publish-npm, release-please]
-    # TODO: Re-enable once Bun SDK compatibility is fixed
-    if: false
-    permissions:
-      contents: write
-
-    steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
-
-      - name: Download all artifacts
-        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c
-        with:
-          path: artifacts
-          merge-multiple: true
-
-      - name: Upload artifacts
-        shell: bash
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          gh release upload \
-            "${{ needs.release-please.outputs.tag_name }}" \
-            artifacts/* \
-            --clobber

.release-please-manifest.json

@@ -1,3 +1,3 @@
 {
-  ".": "0.27.0"
+  ".": "0.30.0"
 }

CHANGELOG.md

@@ -1,5 +1,42 @@
 # Changelog
 
+## [0.30.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.29.2...v0.30.0) (2026-04-20)
+
+
+### Features
+
+* Update to claude-agent-sdk 0.2.114 ([#572](https://github.com/agentclientprotocol/claude-agent-acp/issues/572)) ([e9dd452](https://github.com/agentclientprotocol/claude-agent-acp/commit/e9dd45207d396302f20de82561135bf9558c0e76))
+
+## [0.29.2](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.29.1...v0.29.2) (2026-04-17)
+
+
+### Bug Fixes
+
+* Guard against null usage tokens ([#565](https://github.com/agentclientprotocol/claude-agent-acp/issues/565)) ([f7dc300](https://github.com/agentclientprotocol/claude-agent-acp/commit/f7dc300a165e70f6426f53920f12e24c04033cdc)), closes [#564](https://github.com/agentclientprotocol/claude-agent-acp/issues/564)
+
+## [0.29.1](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.29.0...v0.29.1) (2026-04-17)
+
+
+### Bug Fixes
+
+* emit usage updates from Claude stream events ([#506](https://github.com/agentclientprotocol/claude-agent-acp/issues/506)) ([dd67450](https://github.com/agentclientprotocol/claude-agent-acp/commit/dd67450fd9bddc82db3e71273b535e07b2672804))
+* Remove dot from auto mode description ([#561](https://github.com/agentclientprotocol/claude-agent-acp/issues/561)) ([2ecfa83](https://github.com/agentclientprotocol/claude-agent-acp/commit/2ecfa83b26db58deaded210463fa6ff21d0dff70))
+* Update to claude-agent-sdk 0.2.112 to fix Auto bug with Opus 4.7 ([#562](https://github.com/agentclientprotocol/claude-agent-acp/issues/562)) ([079614a](https://github.com/agentclientprotocol/claude-agent-acp/commit/079614ab05afba17e2cce0d3d238df7b90e17389))
+
+## [0.29.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.28.0...v0.29.0) (2026-04-16)
+
+
+### Features
+
+* Update to claude-agent-sdk 0.2.111 (Opus 4.7) ([#557](https://github.com/agentclientprotocol/claude-agent-acp/issues/557)) ([85cd70c](https://github.com/agentclientprotocol/claude-agent-acp/commit/85cd70c9f3be47c9c404958547c4046d866db1c9))
+
+## [0.28.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.27.0...v0.28.0) (2026-04-15)
+
+
+### Features
+
+* Update to claude-agent-sdk 0.2.109 ([#549](https://github.com/agentclientprotocol/claude-agent-acp/issues/549)) ([07a0fbc](https://github.com/agentclientprotocol/claude-agent-acp/commit/07a0fbc2f6bc388541d064a436412bdd850772cb))
+
 ## [0.27.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.26.0...v0.27.0) (2026-04-13)
 
 

package.json

@@ -3,7 +3,7 @@
   "publishConfig": {
     "access": "public"
   },
-  "version": "0.27.0",
+  "version": "0.30.0",
   "description": "An ACP-compatible coding agent powered by the Claude Agent SDK (TypeScript)",
   "main": "dist/lib.js",
   "types": "dist/lib.d.ts",
@@ -60,22 +60,22 @@
   "author": "Zed Industries",
   "license": "Apache-2.0",
   "dependencies": {
-    "@agentclientprotocol/sdk": "0.18.2",
-    "@anthropic-ai/claude-agent-sdk": "0.2.104",
+    "@agentclientprotocol/sdk": "0.19.0",
+    "@anthropic-ai/claude-agent-sdk": "0.2.114",
     "zod": "^3.25.0 || ^4.0.0"
   },
   "devDependencies": {
-    "@anthropic-ai/sdk": "0.88.0",
+    "@anthropic-ai/sdk": "0.89.0",
     "@eslint/js": "10.0.1",
     "@types/node": "25.6.0",
-    "@typescript-eslint/eslint-plugin": "8.58.1",
-    "@typescript-eslint/parser": "8.58.1",
-    "eslint": "10.2.0",
+    "@typescript-eslint/eslint-plugin": "8.58.2",
+    "@typescript-eslint/parser": "8.58.2",
+    "eslint": "10.2.1",
     "eslint-config-prettier": "10.1.8",
     "globals": "17.5.0",
-    "prettier": "3.8.2",
+    "prettier": "3.8.3",
     "ts-node": "10.9.2",
-    "typescript": "6.0.2",
+    "typescript": "6.0.3",
     "vitest": "4.1.4"
   }
 }

src/acp-agent.ts

@@ -54,7 +54,6 @@ import {
   Query,
   query,
   SDKPartialAssistantMessage,
-  SDKResultMessage,
   SDKUserMessage,
   SlashCommand,
 } from "@anthropic-ai/claude-agent-sdk";
@@ -108,6 +107,22 @@ type AccumulatedUsage = {
   cachedWriteTokens: number;
 };
 
+type UsageSnapshot = {
+  input_tokens: number;
+  output_tokens: number;
+  cache_read_input_tokens: number;
+  cache_creation_input_tokens: number;
+};
+
+const ZERO_USAGE = Object.freeze({
+  input_tokens: 0,
+  output_tokens: 0,
+  cache_read_input_tokens: 0,
+  cache_creation_input_tokens: 0,
+});
+
+const DEFAULT_CONTEXT_WINDOW = 200000;
+
 type Session = {
   query: Query;
   input: Pushable<SDKUserMessage>;
@@ -126,6 +141,12 @@ type Session = {
   nextPendingOrder: number;
   abortController: AbortController;
   emitRawSDKMessages: boolean | SDKMessageFilter[];
+  /** Context window size of the last top-level assistant model, carried across
+   *  prompts so mid-stream usage_update notifications report a correct `size`
+   *  before the turn's first result message arrives. Defaults to
+   *  DEFAULT_CONTEXT_WINDOW, refreshed from each result's modelUsage, and
+   *  invalidated when the user switches the session's model. */
+  contextWindowSize: number;
 };
 
 /** Compute a stable fingerprint of the session-defining params so we can
@@ -239,14 +260,34 @@ export type ToolUseCache = {
   };
 };
 
-function isStaticBinary(): boolean {
-  return process.env.CLAUDE_AGENT_ACP_IS_SINGLE_FILE_BUN !== undefined;
-}
-
 export async function claudeCliPath(): Promise<string> {
-  return isStaticBinary()
-    ? (await import("@anthropic-ai/claude-agent-sdk/embed")).default
-    : import.meta.resolve("@anthropic-ai/claude-agent-sdk").replace("sdk.mjs", "cli.js");
+  if (process.env.CLAUDE_CODE_EXECUTABLE) {
+    return process.env.CLAUDE_CODE_EXECUTABLE;
+  }
+  // The SDK's CLI is a native binary shipped as a platform-specific optional
+  // dependency of @anthropic-ai/claude-agent-sdk. Resolve via a require bound
+  // to the SDK so nested installs are found even when npm doesn't hoist.
+  const { createRequire } = await import("node:module");
+  const req = createRequire(import.meta.resolve("@anthropic-ai/claude-agent-sdk"));
+  const ext = process.platform === "win32" ? ".exe" : "";
+  const candidates =
+    process.platform === "linux"
+      ? [
+          `@anthropic-ai/claude-agent-sdk-linux-${process.arch}-musl/claude${ext}`,
+          `@anthropic-ai/claude-agent-sdk-linux-${process.arch}/claude${ext}`,
+        ]
+      : [`@anthropic-ai/claude-agent-sdk-${process.platform}-${process.arch}/claude${ext}`];
+  for (const candidate of candidates) {
+    try {
+      return req.resolve(candidate);
+    } catch {
+      // try next candidate
+    }
+  }
+  throw new Error(
+    `Claude native binary not found for ${process.platform}-${process.arch}. ` +
+      `Reinstall @anthropic-ai/claude-agent-sdk without --omit=optional, or set CLAUDE_CODE_EXECUTABLE.`,
+  );
 }
 
 function shouldHideClaudeAuth(): boolean {
@@ -545,8 +586,8 @@ export class ClaudeAcpAgent implements Agent {
     };
 
     let lastAssistantTotalUsage: number | null = null;
+    let lastAssistantUsage: UsageSnapshot | null = null;
     let lastAssistantModel: string | null = null;
-    let lastContextWindowSize: number = 200000;
 
     const userMessage = promptToClaude(params);
 
@@ -628,12 +669,13 @@ export class ClaudeAcpAgent implements Agent {
                 // "944k/1m" right after the user sees "Compacting completed",
                 // which is confusing and wrong.
                 lastAssistantTotalUsage = 0;
+                lastAssistantUsage = null;
                 await this.client.sessionUpdate({
                   sessionId: message.session_id,
                   update: {
                     sessionUpdate: "usage_update",
                     used: 0,
-                    size: lastContextWindowSize,
+                    size: session.contextWindowSize,
                   },
                 });
                 await this.client.sessionUpdate({
@@ -670,7 +712,11 @@ export class ClaudeAcpAgent implements Agent {
               case "task_progress":
               case "task_updated":
               case "elicitation_complete":
+              case "plugin_install":
+              case "memory_recall":
+              case "notification":
               case "api_retry":
+              case "mirror_error":
                 // Todo: process via status api: https://docs.claude.com/en/docs/claude-code/hooks#hook-output
                 break;
               default:
@@ -688,8 +734,13 @@ export class ClaudeAcpAgent implements Agent {
             const matchingModelUsage = lastAssistantModel
               ? getMatchingModelUsage(message.modelUsage, lastAssistantModel)
               : null;
-            const contextWindowSize = matchingModelUsage?.contextWindow ?? 200000;
-            lastContextWindowSize = contextWindowSize;
+            // Only overwrite when we have an authoritative value — a miss
+            // (e.g. a turn with no top-level assistant message) would
+            // otherwise discard the window learned on a prior turn and
+            // leave the next prompt's mid-stream updates reporting 200k.
+            if (matchingModelUsage) {
+              session.contextWindowSize = matchingModelUsage.contextWindow;
+            }
 
             // Send usage_update notification
             if (lastAssistantTotalUsage !== null) {
@@ -698,7 +749,7 @@ export class ClaudeAcpAgent implements Agent {
                 update: {
                   sessionUpdate: "usage_update",
                   used: lastAssistantTotalUsage,
-                  size: contextWindowSize,
+                  size: session.contextWindowSize,
                   cost: {
                     amount: message.total_cost_usd,
                     currency: "USD",
@@ -772,6 +823,57 @@ export class ClaudeAcpAgent implements Agent {
             break;
           }
           case "stream_event": {
+            if (
+              message.parent_tool_use_id === null &&
+              (message.event.type === "message_start" || message.event.type === "message_delta")
+            ) {
+              if (message.event.type === "message_start") {
+                lastAssistantUsage = snapshotFromUsage(message.event.message.usage);
+                const model = message.event.message.model;
+                if (model && model !== "<synthetic>") {
+                  lastAssistantModel = model;
+                  // Only upgrade from the default — once a `result` has given
+                  // us an authoritative window, trust it over the heuristic.
+                  // Model switches invalidate the cached window via
+                  // `syncSessionConfigState`, which resets us back to the
+                  // default so this branch runs again for the new model.
+                  if (session.contextWindowSize === DEFAULT_CONTEXT_WINDOW) {
+                    const inferred = inferContextWindowFromModel(model);
+                    if (inferred !== null) {
+                      session.contextWindowSize = inferred;
+                    }
+                  }
+                }
+              } else {
+                const usage = message.event.usage;
+                const prev: Readonly<UsageSnapshot> = lastAssistantUsage ?? ZERO_USAGE;
+                // Per Anthropic API, message_delta usage fields are *cumulative*;
+                // nullable fields (input_tokens and the cache fields) fall back
+                // to the prior snapshot when the server omits them from this
+                // delta. Only output_tokens is guaranteed non-null.
+                lastAssistantUsage = {
+                  input_tokens: usage.input_tokens ?? prev.input_tokens,
+                  output_tokens: usage.output_tokens,
+                  cache_read_input_tokens:
+                    usage.cache_read_input_tokens ?? prev.cache_read_input_tokens,
+                  cache_creation_input_tokens:
+                    usage.cache_creation_input_tokens ?? prev.cache_creation_input_tokens,
+                };
+              }
+
+              const nextUsage = totalTokens(lastAssistantUsage);
+              if (nextUsage !== lastAssistantTotalUsage) {
+                lastAssistantTotalUsage = nextUsage;
+                await this.client.sessionUpdate({
+                  sessionId: params.sessionId,
+                  update: {
+                    sessionUpdate: "usage_update",
+                    used: nextUsage,
+                    size: session.contextWindowSize,
+                  },
+                });
+              }
+            }
             for (const notification of streamEventToAcpNotifications(
               message,
               params.sessionId,
@@ -814,29 +916,16 @@ export class ClaudeAcpAgent implements Agent {
               }
             }
 
-            // Store latest assistant usage (excluding subagents)
-            // Sum all token types as a proxy for post-turn context occupancy:
-            // current turn's output will become next turn's input.
-            // Note: per the Anthropic API, input_tokens excludes cache tokens —
-            // cache_read and cache_creation are reported separately, so summing
-            // all four fields is not double-counting.
-            if ((message.message as any).usage && message.parent_tool_use_id === null) {
-              const messageWithUsage = message.message as unknown as SDKResultMessage;
-              lastAssistantTotalUsage =
-                messageWithUsage.usage.input_tokens +
-                messageWithUsage.usage.output_tokens +
-                messageWithUsage.usage.cache_read_input_tokens +
-                messageWithUsage.usage.cache_creation_input_tokens;
-            }
-            // Track the current top-level model for context window size lookup
-            // (exclude subagent messages to stay in sync with lastAssistantTotalUsage)
-            if (
-              message.type === "assistant" &&
-              message.parent_tool_use_id === null &&
-              message.message.model &&
-              message.message.model !== "<synthetic>"
-            ) {
-              lastAssistantModel = message.message.model;
+            // Snapshot the latest top-level assistant usage and model so the
+            // next `result` can emit a usage_update tied to the right context
+            // window. Subagent messages are excluded to keep the snapshot
+            // aligned with what the user's current selection is producing.
+            if (message.type === "assistant" && message.parent_tool_use_id === null) {
+              lastAssistantUsage = snapshotFromUsage(message.message.usage);
+              lastAssistantTotalUsage = totalTokens(lastAssistantUsage);
+              if (message.message.model && message.message.model !== "<synthetic>") {
+                lastAssistantModel = message.message.model;
+              }
             }
 
             // Slash commands like /compact can generate invalid output... doesn't match
@@ -1331,6 +1420,13 @@ export class ClaudeAcpAgent implements Agent {
     if (configId === "mode") {
       session.modes = { ...session.modes, currentModeId: value };
     } else if (configId === "model") {
+      if (session.models.currentModelId !== value) {
+        // The cached context window was learned for the previous model; reset
+        // to the new model's heuristic so mid-stream updates between now and
+        // the next `result` reflect the user's selection instead of the old
+        // model's window.
+        session.contextWindowSize = inferContextWindowFromModel(value) ?? DEFAULT_CONTEXT_WINDOW;
+      }
       session.models = { ...session.models, currentModelId: value };
     }
   }
@@ -1487,14 +1583,7 @@ export class ClaudeAcpAgent implements Agent {
       allowDangerouslySkipPermissions: ALLOW_BYPASS,
       permissionMode,
       canUseTool: this.canUseTool(sessionId),
-      // note: although not documented by the types, passing an absolute path
-      // here works to find zed's managed node version.
-      executable: isStaticBinary() ? undefined : (process.execPath as any),
-      ...(process.env.CLAUDE_CODE_EXECUTABLE
-        ? { pathToClaudeCodeExecutable: process.env.CLAUDE_CODE_EXECUTABLE }
-        : isStaticBinary()
-          ? { pathToClaudeCodeExecutable: await claudeCliPath() }
-          : {}),
+      pathToClaudeCodeExecutable: process.env.CLAUDE_CODE_EXECUTABLE,
       extraArgs: {
         ...userProvidedOptions?.extraArgs,
         "replay-user-messages": "",
@@ -1554,7 +1643,8 @@ export class ClaudeAcpAgent implements Agent {
       if (
         creationOpts.resume &&
         error instanceof Error &&
-        error.message === "Query closed before response received"
+        (error.message === "Query closed before response received" ||
+          error.message.includes("No conversation found with session ID"))
       ) {
         throw RequestError.resourceNotFound(sessionId);
       }
@@ -1578,7 +1668,7 @@ export class ClaudeAcpAgent implements Agent {
       {
         id: "auto",
         name: "Auto",
-        description: "Use a model classifier to approve/deny permission prompts.",
+        description: "Use a model classifier to approve/deny permission prompts",
       },
       {
         id: "default",
@@ -1638,6 +1728,8 @@ export class ClaudeAcpAgent implements Agent {
       nextPendingOrder: 0,
       abortController,
       emitRawSDKMessages: sessionMeta?.claudeCode?.emitRawSDKMessages ?? false,
+      contextWindowSize:
+        inferContextWindowFromModel(models.currentModelId) ?? DEFAULT_CONTEXT_WINDOW,
     };
 
     return {
@@ -1674,6 +1766,41 @@ function sessionUsage(session: Session) {
   };
 }
 
+/** Sum all four fields as a proxy for post-turn context occupancy: the current
+ *  turn's output becomes next turn's input. Per the Anthropic API, input_tokens
+ *  excludes cache tokens — cache_read and cache_creation are reported
+ *  separately — so summing all four is not double-counting. */
+function totalTokens(usage: UsageSnapshot): number {
+  return (
+    usage.input_tokens +
+    usage.output_tokens +
+    usage.cache_read_input_tokens +
+    usage.cache_creation_input_tokens
+  );
+}
+
+/** Project a nullable API usage object into our non-null snapshot shape.
+ *  Both SDK message_start and assistant message `usage` have `number | null`
+ *  cache fields; we coerce absent values to 0 so `totalTokens` never hits
+ *  NaN. `input_tokens`/`output_tokens` are typed `number` by the SDK but
+ *  synthetic or third-party-backend stream events have been observed emitting
+ *  them as null/undefined — coerce those too so a malformed upstream event
+ *  can't leak NaN into the wire `used` field. Delta events have different
+ *  semantics (cumulative + prev fallback) and are handled inline. */
+function snapshotFromUsage(usage: {
+  input_tokens?: number | null;
+  output_tokens?: number | null;
+  cache_read_input_tokens?: number | null;
+  cache_creation_input_tokens?: number | null;
+}): UsageSnapshot {
+  return {
+    input_tokens: usage.input_tokens ?? 0,
+    output_tokens: usage.output_tokens ?? 0,
+    cache_read_input_tokens: usage.cache_read_input_tokens ?? 0,
+    cache_creation_input_tokens: usage.cache_creation_input_tokens ?? 0,
+  };
+}
+
 function createEnvForGateway(gatewayMeta?: GatewayAuthMeta) {
   if (!gatewayMeta) {
     return {};
@@ -2298,6 +2425,16 @@ function commonPrefixLength(a: string, b: string) {
   return i;
 }
 
+/** Best-effort first guess of a model's context window from its ID, used only
+ *  until a `result` message arrives with the authoritative `modelUsage` value.
+ *  Anthropic 1M-context variants encode "1m" as a distinct token in the SDK
+ *  model ID (e.g., "claude-opus-4-6-1m"), which `\b1m\b` catches without also
+ *  matching things like "10m" or embedded substrings. */
+function inferContextWindowFromModel(model: string): number | null {
+  if (/\b1m\b/i.test(model)) return 1_000_000;
+  return null;
+}
+
 function getMatchingModelUsage(modelUsage: Record<string, ModelUsage>, currentModel: string) {
   let bestKey: string | null = null;
   let bestLen = 0;

src/index.ts

@@ -5,8 +5,35 @@ import { loadManagedSettings, applyEnvironmentSettings } from "./utils.js";
 import { claudeCliPath, runAcp } from "./acp-agent.js";
 
 if (process.argv.includes("--cli")) {
-  process.argv = process.argv.filter((arg) => arg !== "--cli");
-  await import(await claudeCliPath());
+  const { spawn } = await import("node:child_process");
+  const args = process.argv.slice(2).filter((arg) => arg !== "--cli");
+  const child = spawn(await claudeCliPath(), args, { stdio: "inherit" });
+
+  const signals =
+    process.platform === "win32"
+      ? (["SIGINT", "SIGTERM"] as const)
+      : (["SIGINT", "SIGTERM", "SIGHUP"] as const);
+  for (const sig of signals) {
+    process.on(sig, () => {
+      if (!child.killed) child.kill(sig);
+    });
+  }
+
+  child.on("exit", (code, signal) => {
+    if (signal && process.platform !== "win32") {
+      // Remove our listener so re-raising actually terminates instead of
+      // re-entering the no-op handler, which would let us exit with code 0
+      // instead of the signal's conventional 128+N.
+      process.removeAllListeners(signal);
+      process.kill(process.pid, signal);
+    } else {
+      process.exit(code ?? 1);
+    }
+  });
+  child.on("error", (err) => {
+    console.error(err);
+    process.exit(1);
+  });
 } else {
   const managedSettings = loadManagedSettings();
   if (managedSettings) {

src/tests/acp-agent.test.ts

@@ -214,8 +214,7 @@ describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("ACP subprocess integration"
     const commands = await client.availableCommandsPromise;
 
     expect(commands).toContainEqual({
-      description:
-        "Clear conversation history but keep a summary in context. Optional: /compact [instructions for summarization]",
+      description: "Free up context by summarizing the conversation so far",
       input: {
         hint: "<optional custom summarization instructions>",
       },
@@ -1173,7 +1172,7 @@ describe("prompt conversion", () => {
 describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("SDK behavior", () => {
   it("finds vendored cli path", async () => {
     const path = await claudeCliPath();
-    expect(path).toContain("@anthropic-ai/claude-agent-sdk/cli.js");
+    expect(path).toMatch(/@anthropic-ai\/claude-agent-sdk-[^/]+\/claude(\.exe)?$/);
   });
 
   it("query has a 'default' model", async () => {
@@ -1350,6 +1349,7 @@ describe("stop reason propagation", () => {
       nextPendingOrder: 0,
       abortController: new AbortController(),
       emitRawSDKMessages: false,
+      contextWindowSize: 200000,
     };
   }
 
@@ -1491,6 +1491,7 @@ describe("stop reason propagation", () => {
       pendingMessages: new Map(),
       nextPendingOrder: 0,
       emitRawSDKMessages: false,
+      contextWindowSize: 200000,
     };
 
     const response = await agent.prompt({
@@ -1566,6 +1567,7 @@ describe("session/close", () => {
       nextPendingOrder: 0,
       abortController: new AbortController(),
       emitRawSDKMessages: false,
+      contextWindowSize: 200000,
     };
     return agent.sessions[sessionId]!;
   }
@@ -1660,6 +1662,7 @@ describe("getOrCreateSession param change detection", () => {
       nextPendingOrder: 0,
       abortController: new AbortController(),
       emitRawSDKMessages: false,
+      contextWindowSize: 200000,
     };
     return agent.sessions[sessionId]!;
   }
@@ -1819,6 +1822,23 @@ describe("usage_update computation", () => {
     };
   }
 
+  function createStreamEvent(
+    eventType: "message_start" | "message_delta",
+    payload: Record<string, unknown>,
+    parentToolUseId: string | null = null,
+  ) {
+    return {
+      type: "stream_event" as const,
+      parent_tool_use_id: parentToolUseId,
+      uuid: randomUUID(),
+      session_id: "test-session",
+      event:
+        eventType === "message_start"
+          ? { type: "message_start" as const, message: payload }
+          : { type: "message_delta" as const, ...payload },
+    };
+  }
+
   function createMockAgentWithCapture() {
     const updates: any[] = [];
     const mockClient = {
@@ -1875,6 +1895,7 @@ describe("usage_update computation", () => {
       nextPendingOrder: 0,
       abortController: new AbortController(),
       emitRawSDKMessages: false,
+      contextWindowSize: 200000,
     };
   }
 
@@ -1915,6 +1936,513 @@ describe("usage_update computation", () => {
     expect(usageUpdate.update.used).toBe(1800);
   });
 
+  it("coerces null input/output tokens so wire `used` is never null", async () => {
+    // Synthetic or third-party-backend stream events have been observed
+    // emitting input_tokens/output_tokens as null. Without coercion the
+    // snapshot leaks NaN into totalTokens(), and JSON.stringify(NaN) === "null"
+    // produces a malformed `used: null` that schema-validating ACP clients reject.
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createAssistantMessage({
+        model: "claude-opus-4-20250514",
+        usage: {
+          input_tokens: null,
+          output_tokens: null,
+          cache_read_input_tokens: 200,
+          cache_creation_input_tokens: 100,
+        } as unknown as {
+          input_tokens: number;
+          output_tokens: number;
+          cache_read_input_tokens: number;
+          cache_creation_input_tokens: number;
+        },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-20250514": {
+            inputTokens: 0,
+            outputTokens: 0,
+            cacheReadInputTokens: 200,
+            cacheCreationInputTokens: 100,
+            webSearchRequests: 0,
+            costUSD: 0.01,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates.length).toBeGreaterThan(0);
+    for (const u of usageUpdates) {
+      expect(u.update.used).not.toBeNull();
+      expect(Number.isFinite(u.update.used)).toBe(true);
+      // Round-trip through JSON to catch the NaN -> "null" serialization bug.
+      const wire = JSON.parse(JSON.stringify(u.update));
+      expect(wire.used).not.toBeNull();
+      expect(typeof wire.used).toBe("number");
+    }
+  });
+
+  it("stream_event message_start emits usage_update before result", async () => {
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-opus-4-20250514",
+        usage: {
+          input_tokens: 1000,
+          output_tokens: 500,
+          cache_read_input_tokens: 200,
+          cache_creation_input_tokens: 100,
+        },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-20250514": {
+            inputTokens: 1000,
+            outputTokens: 500,
+            cacheReadInputTokens: 200,
+            cacheCreationInputTokens: 100,
+            webSearchRequests: 0,
+            costUSD: 0.01,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(2);
+    expect(usageUpdates[0].update.used).toBe(1800);
+    // First prompt of a session has no prior result to learn the window from,
+    // so the mid-stream update falls back to the default context window.
+    expect(usageUpdates[0].update.size).toBe(200000);
+    expect(usageUpdates[0].update.cost).toBeUndefined();
+    expect(usageUpdates[1].update.used).toBe(1800);
+    expect(usageUpdates[1].update.size).toBe(1000000);
+    expect(usageUpdates[1].update.cost).toBeDefined();
+  });
+
+  it("stream_event message_delta patches previous snapshot", async () => {
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-opus-4-20250514",
+        usage: {
+          input_tokens: 1000,
+          output_tokens: 0,
+          cache_read_input_tokens: 200,
+          cache_creation_input_tokens: 100,
+        },
+      }),
+      createStreamEvent("message_delta", {
+        usage: { output_tokens: 500 },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-20250514": {
+            inputTokens: 1000,
+            outputTokens: 500,
+            cacheReadInputTokens: 200,
+            cacheCreationInputTokens: 100,
+            webSearchRequests: 0,
+            costUSD: 0.01,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(3);
+    expect(usageUpdates[0].update.used).toBe(1300);
+    expect(usageUpdates[0].update.cost).toBeUndefined();
+    expect(usageUpdates[1].update.used).toBe(1800);
+    expect(usageUpdates[1].update.cost).toBeUndefined();
+    expect(usageUpdates[2].update.used).toBe(1800);
+    expect(usageUpdates[2].update.cost).toBeDefined();
+  });
+
+  it("mid-stream size is inferred from a 1M model name before the first result", async () => {
+    // On the very first prompt there is no learned context window yet, so the
+    // mid-stream update would otherwise fall back to 200k. A "-1m" suffix in
+    // the SDK model ID is enough signal to emit 1_000_000 up front.
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-opus-4-6-1m",
+        usage: {
+          input_tokens: 2000,
+          output_tokens: 1000,
+          cache_read_input_tokens: 0,
+          cache_creation_input_tokens: 0,
+        },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-6-1m": {
+            inputTokens: 2000,
+            outputTokens: 1000,
+            cacheReadInputTokens: 0,
+            cacheCreationInputTokens: 0,
+            webSearchRequests: 0,
+            costUSD: 0.02,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(2);
+    expect(usageUpdates[0].update.size).toBe(1000000);
+    expect(usageUpdates[1].update.size).toBe(1000000);
+  });
+
+  it("duplicate stream_event totals do not re-emit usage_update", async () => {
+    // A message_delta whose cumulative totals match the prior snapshot should
+    // not trigger a duplicate usage_update — only the result adds cost on top.
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-opus-4-20250514",
+        usage: {
+          input_tokens: 1000,
+          output_tokens: 500,
+          cache_read_input_tokens: 200,
+          cache_creation_input_tokens: 100,
+        },
+      }),
+      createStreamEvent("message_delta", {
+        usage: { output_tokens: 500 },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-20250514": {
+            inputTokens: 1000,
+            outputTokens: 500,
+            cacheReadInputTokens: 200,
+            cacheCreationInputTokens: 100,
+            webSearchRequests: 0,
+            costUSD: 0.01,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(2);
+    expect(usageUpdates[0].update.used).toBe(1800);
+    expect(usageUpdates[0].update.cost).toBeUndefined();
+    expect(usageUpdates[1].update.used).toBe(1800);
+    expect(usageUpdates[1].update.cost).toBeDefined();
+  });
+
+  it("mid-stream size uses the session's learned context window", async () => {
+    // Session state persists the model's context window across prompts, so a
+    // mid-stream update in a later prompt reports the real size immediately
+    // instead of snapping back to the 200k default before the result arrives.
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-opus-4-20250514",
+        usage: {
+          input_tokens: 2000,
+          output_tokens: 1000,
+          cache_read_input_tokens: 0,
+          cache_creation_input_tokens: 0,
+        },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-20250514": {
+            inputTokens: 2000,
+            outputTokens: 1000,
+            cacheReadInputTokens: 0,
+            cacheCreationInputTokens: 0,
+            webSearchRequests: 0,
+            costUSD: 0.02,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+    // Simulate a prior prompt having learned the 1M window for this model.
+    agent.sessions["test-session"].contextWindowSize = 1000000;
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(2);
+    expect(usageUpdates[0].update.size).toBe(1000000);
+    expect(usageUpdates[1].update.size).toBe(1000000);
+  });
+
+  it("switching to a 1M model seeds the context window from the heuristic", async () => {
+    // The heuristic runs at config-change time so mid-stream updates in the
+    // next prompt already report 1M — without waiting for message_start or
+    // the next `result` to correct us.
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-opus-4-6-1m",
+        usage: {
+          input_tokens: 2000,
+          output_tokens: 1000,
+          cache_read_input_tokens: 0,
+          cache_creation_input_tokens: 0,
+        },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-6-1m": {
+            inputTokens: 2000,
+            outputTokens: 1000,
+            cacheReadInputTokens: 0,
+            cacheCreationInputTokens: 0,
+            webSearchRequests: 0,
+            costUSD: 0.02,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+    const session = agent.sessions["test-session"];
+    expect(session.contextWindowSize).toBe(200000);
+
+    (agent as any).syncSessionConfigState(session, "model", "claude-opus-4-6-1m");
+    expect(session.contextWindowSize).toBe(1000000);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(2);
+    expect(usageUpdates[0].update.size).toBe(1000000);
+    expect(usageUpdates[1].update.size).toBe(1000000);
+  });
+
+  it("result with no matching modelUsage preserves the learned window", async () => {
+    // A turn whose `result.modelUsage` doesn't contain the current top-level
+    // model (e.g. no top-level assistant message, or only a subagent ran) must
+    // not clobber the window learned on a prior turn — otherwise the next
+    // prompt's mid-stream updates regress to the 200k default.
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-haiku-4-5-20251001": {
+            inputTokens: 100,
+            outputTokens: 50,
+            cacheReadInputTokens: 0,
+            cacheCreationInputTokens: 0,
+            webSearchRequests: 0,
+            costUSD: 0.001,
+            contextWindow: 200000,
+            maxOutputTokens: 8192,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+    const session = agent.sessions["test-session"];
+    session.contextWindowSize = 1000000;
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    expect(session.contextWindowSize).toBe(1000000);
+    // The emit itself falls back to session.contextWindowSize, which is
+    // unchanged from the learned value.
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    // No lastAssistantTotalUsage was set (no top-level assistant / stream
+    // event), so the result branch skips its emit entirely.
+    expect(usageUpdates).toHaveLength(0);
+  });
+
+  it("switching the session's model invalidates the learned context window", async () => {
+    // When the user switches models mid-session, the window learned for the
+    // previous model would otherwise persist into the next prompt's first
+    // mid-stream update. syncSessionConfigState should reset it so the next
+    // turn's first update falls back to the heuristic (here: 200k default).
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-sonnet-4-6",
+        usage: {
+          input_tokens: 1000,
+          output_tokens: 500,
+          cache_read_input_tokens: 0,
+          cache_creation_input_tokens: 0,
+        },
+      }),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-sonnet-4-6": {
+            inputTokens: 1000,
+            outputTokens: 500,
+            cacheReadInputTokens: 0,
+            cacheCreationInputTokens: 0,
+            webSearchRequests: 0,
+            costUSD: 0.01,
+            contextWindow: 200000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+    const session = agent.sessions["test-session"];
+    session.contextWindowSize = 1000000;
+    session.models = { ...session.models, currentModelId: "claude-opus-4-6-1m" };
+
+    // User flips the selector to a 200k model.
+    (agent as any).syncSessionConfigState(session, "model", "claude-sonnet-4-6");
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(2);
+    expect(usageUpdates[0].update.size).toBe(200000);
+    expect(usageUpdates[1].update.size).toBe(200000);
+  });
+
+  it("non-usage stream events do not re-emit usage_update", async () => {
+    // content_block_* and message_stop carry no usage fields; they must not
+    // trigger duplicate emits between the real message_start / message_delta
+    // / result updates.
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent("message_start", {
+        model: "claude-opus-4-20250514",
+        usage: {
+          input_tokens: 1000,
+          output_tokens: 0,
+          cache_read_input_tokens: 0,
+          cache_creation_input_tokens: 0,
+        },
+      }),
+      {
+        type: "stream_event" as const,
+        parent_tool_use_id: null,
+        uuid: randomUUID(),
+        session_id: "test-session",
+        event: { type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
+      },
+      {
+        type: "stream_event" as const,
+        parent_tool_use_id: null,
+        uuid: randomUUID(),
+        session_id: "test-session",
+        event: { type: "content_block_delta", index: 0, delta: { type: "text_delta", text: "hi" } },
+      },
+      {
+        type: "stream_event" as const,
+        parent_tool_use_id: null,
+        uuid: randomUUID(),
+        session_id: "test-session",
+        event: { type: "content_block_stop", index: 0 },
+      },
+      createStreamEvent("message_delta", {
+        usage: { output_tokens: 200 },
+      }),
+      {
+        type: "stream_event" as const,
+        parent_tool_use_id: null,
+        uuid: randomUUID(),
+        session_id: "test-session",
+        event: { type: "message_stop" },
+      },
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-opus-4-20250514": {
+            inputTokens: 1000,
+            outputTokens: 200,
+            cacheReadInputTokens: 0,
+            cacheCreationInputTokens: 0,
+            webSearchRequests: 0,
+            costUSD: 0.01,
+            contextWindow: 1000000,
+            maxOutputTokens: 16384,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    // Exactly three: message_start (1000), message_delta (1200), result (1200 + cost).
+    expect(usageUpdates).toHaveLength(3);
+    expect(usageUpdates[0].update.used).toBe(1000);
+    expect(usageUpdates[1].update.used).toBe(1200);
+    expect(usageUpdates[2].update.used).toBe(1200);
+    expect(usageUpdates[2].update.cost).toBeDefined();
+  });
+
+  it("subagent stream_event does not emit usage_update", async () => {
+    const { agent, updates } = createMockAgentWithCapture();
+    injectSession(agent, [
+      createStreamEvent(
+        "message_start",
+        {
+          model: "claude-haiku-4-5-20251001",
+          usage: {
+            input_tokens: 500,
+            output_tokens: 100,
+            cache_read_input_tokens: 0,
+            cache_creation_input_tokens: 0,
+          },
+        },
+        "tool_use_123",
+      ),
+      createResultMessageWithModel({
+        modelUsage: {
+          "claude-haiku-4-5-20251001": {
+            inputTokens: 500,
+            outputTokens: 100,
+            cacheReadInputTokens: 0,
+            cacheCreationInputTokens: 0,
+            webSearchRequests: 0,
+            costUSD: 0.001,
+            contextWindow: 200000,
+            maxOutputTokens: 8192,
+          },
+        },
+      }),
+      { type: "system", subtype: "session_state_changed", state: "idle" },
+    ]);
+
+    await agent.prompt({ sessionId: "test-session", prompt: [{ type: "text", text: "test" }] });
+
+    const usageUpdates = updates.filter((u: any) => u.update?.sessionUpdate === "usage_update");
+    expect(usageUpdates).toHaveLength(0);
+  });
+
   it("size reflects the current model's context window, not min across all", async () => {
     const { agent, updates } = createMockAgentWithCapture();
     injectSession(agent, [
@@ -2255,6 +2783,7 @@ describe("emitRawSDKMessages", () => {
       nextPendingOrder: 0,
       abortController: new AbortController(),
       emitRawSDKMessages,
+      contextWindowSize: 200000,
     };
   }
 

src/tests/session-config-options.test.ts

@@ -93,7 +93,10 @@ describe("session config options", () => {
       cancelled: false,
       permissionMode: "default",
       settingsManager: {},
+      modes: structuredClone(MOCK_MODES),
+      models: structuredClone(MOCK_MODELS),
       configOptions: structuredClone(MOCK_CONFIG_OPTIONS),
+      contextWindowSize: 200000,
     };
   }
 

src/tests/session-load.test.ts

@@ -0,0 +1,148 @@
+import { describe, it, expect } from "vitest";
+import { randomUUID } from "crypto";
+import { AgentSideConnection, RequestError, SessionNotification } from "@agentclientprotocol/sdk";
+import { query, getSessionMessages } from "@anthropic-ai/claude-agent-sdk";
+import { ClaudeAcpAgent } from "../acp-agent.js";
+import { Pushable } from "../utils.js";
+import type { SDKUserMessage } from "@anthropic-ai/claude-agent-sdk";
+
+function createMockClient(): AgentSideConnection {
+  return {
+    sessionUpdate: async (_notification: SessionNotification) => {},
+    requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
+    readTextFile: async () => ({ content: "" }),
+    writeTextFile: async () => ({}),
+  } as unknown as AgentSideConnection;
+}
+
+describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("session load/resume lifecycle", () => {
+  it("SDK: session created but never prompted has no messages and is not resumable", async () => {
+    // Create a session via the SDK, initialize it, but never send a prompt
+    const sessionId = randomUUID();
+    const input = new Pushable<SDKUserMessage>();
+
+    const q = query({
+      prompt: input,
+      options: {
+        systemPrompt: { type: "preset", preset: "claude_code" },
+        sessionId,
+        settingSources: ["user", "project", "local"],
+        includePartialMessages: true,
+      },
+    });
+
+    // initializationResult() works without needing a prompt pushed
+    const initResult = await q.initializationResult();
+    expect(initResult).toBeDefined();
+
+    // Close without ever prompting
+    input.end();
+    q.return(undefined);
+
+    // Verify no messages were stored
+    const messages = await getSessionMessages(sessionId);
+    expect(messages).toEqual([]);
+
+    // Verify the session is not resumable
+    const input2 = new Pushable<SDKUserMessage>();
+    const q2 = query({
+      prompt: input2,
+      options: {
+        systemPrompt: { type: "preset", preset: "claude_code" },
+        resume: sessionId,
+        settingSources: ["user", "project", "local"],
+        includePartialMessages: true,
+      },
+    });
+
+    await expect(q2.initializationResult()).rejects.toThrow(
+      /No conversation found with session ID/,
+    );
+
+    input2.end();
+    q2.return(undefined);
+  }, 30000);
+
+  it("ACP: loadSession throws resourceNotFound for a non-existent session", async () => {
+    const agent = new ClaudeAcpAgent(createMockClient());
+    const bogusSessionId = randomUUID();
+
+    try {
+      await expect(
+        agent.loadSession({
+          sessionId: bogusSessionId,
+          cwd: process.cwd(),
+          mcpServers: [],
+        }),
+      ).rejects.toThrow(RequestError);
+    } finally {
+      await agent.dispose();
+    }
+  }, 30000);
+
+  it("ACP: resumeSession throws resourceNotFound for a non-existent session", async () => {
+    const agent = new ClaudeAcpAgent(createMockClient());
+    const bogusSessionId = randomUUID();
+
+    try {
+      await expect(
+        agent.unstable_resumeSession({
+          sessionId: bogusSessionId,
+          cwd: process.cwd(),
+          mcpServers: [],
+        }),
+      ).rejects.toThrow(RequestError);
+    } finally {
+      await agent.dispose();
+    }
+  }, 30000);
+
+  it("ACP: newSession without prompt, then loadSession on fresh agent throws resourceNotFound", async () => {
+    // Step 1: Create a real session via ACP, never prompt, dispose
+    const agentA = new ClaudeAcpAgent(createMockClient());
+    const { sessionId } = await agentA.newSession({
+      cwd: process.cwd(),
+      mcpServers: [],
+    });
+    expect(sessionId).toBeDefined();
+    await agentA.dispose();
+
+    // Step 2: Fresh agent tries to load that session
+    const agentB = new ClaudeAcpAgent(createMockClient());
+
+    try {
+      await expect(
+        agentB.loadSession({
+          sessionId,
+          cwd: process.cwd(),
+          mcpServers: [],
+        }),
+      ).rejects.toThrow(RequestError);
+    } finally {
+      await agentB.dispose();
+    }
+  }, 30000);
+
+  it("ACP: newSession without prompt, then resumeSession on fresh agent throws resourceNotFound", async () => {
+    const agentA = new ClaudeAcpAgent(createMockClient());
+    const { sessionId } = await agentA.newSession({
+      cwd: process.cwd(),
+      mcpServers: [],
+    });
+    await agentA.dispose();
+
+    const agentB = new ClaudeAcpAgent(createMockClient());
+
+    try {
+      await expect(
+        agentB.unstable_resumeSession({
+          sessionId,
+          cwd: process.cwd(),
+          mcpServers: [],
+        }),
+      ).rejects.toThrow(RequestError);
+    } finally {
+      await agentB.dispose();
+    }
+  }, 30000);
+});