Hook callback errors (Stream closed) on every tool call since v0.1.29

[sebyx07]

Bug Description

Since upgrading to claude-agent-sdk v0.1.29+, every tool call produces noisy Error in hook callback hook_0: Stream closed errors in the output. The errors are non-fatal (tools complete successfully and execution continues), but they spam the console with ~50 lines of minified Claude CLI source code per tool call, making logs unreadable.

Environment

• claude-agent-sdk: 0.1.31 (also reproduces on 0.1.29 and 0.1.30)
• Python: 3.12
• Platform: Ubuntu 24.04 (Linux 6.14.0)
• Claude CLI bundled version: 2.1.33 (v0.1.31), 2.1.31 (v0.1.29)
• Permission mode: bypassPermissions
• Usage: Programmatic agent via `query()` with `ClaudeAgentOptions(allowed_tools=[...], permission_mode="bypassPermissions")`

Steps to Reproduce

1. Install `claude-agent-sdk>=0.1.29`
2. Run any agent query using `bypassPermissions` mode
3. Observe every tool call (Read, Glob, Grep, Bash, etc.) produces the error

Error Output (per tool call)

[claude 10:58:16] Using tool: Read → CLAUDE.md
Error in hook callback hook_0: 6339 | Claude Code has been suspended...
...~50 lines of minified CLI source code...
6344 | `)}async sendRequest(H,$,A){...if(this.inputClosed)throw Error("Stream closed")...

error: Stream closed
      at sendRequest (/$bunfs/root/claude:6344:133)
      at sendRequest (/$bunfs/root/claude:6344:627)
      at <anonymous> (/$bunfs/root/claude:6344:1305)
      at callback (/$bunfs/root/claude:6344:1451)
      at _U1 (/$bunfs/root/claude:3151:1288)
      at _U1 (/$bunfs/root/claude:3151:1525)
      at <anonymous> (/$bunfs/root/claude:3144:1239)

[claude 10:58:16] Tool completed    ← tool still succeeds

This repeats for every single tool call in a session. A typical planning phase with 20+ tool calls produces hundreds of lines of error output.

Root Cause Analysis

v0.1.29 introduced new hook event types:

• `SubagentStop`
• `Notification`
• `PermissionRequest`

The error appears to originate from the parent process's hook callback trying to handle permission requests via `sendRequest` over a stream that's closed or not applicable when `bypassPermissions` mode is active. The hook fires, fails to communicate over the stream, throws `Stream closed`, but execution continues because permissions are bypassed.

The minified source visible in the error includes the `SubagentStop` hook handling code:

hookEvent==="Stop"||hookEvent==="SubagentStop"

Workaround

Pinning to `claude-agent-sdk<0.1.29` (v0.1.28, CLI v2.1.30) eliminates the errors entirely.

Expected Behavior

Tool calls in `bypassPermissions` mode should not trigger hook callbacks that attempt stream communication, or the errors should be silently caught rather than dumped to stderr with minified source code.

Impact

• Console output is unreadable during agent execution
• Log files become bloated (each tool call adds ~50 lines of noise)
• Makes it difficult to debug actual issues in agent output
• Confusing for users who think the agent is crashing (it's not)

[sebyx07]

My Environment

• OS: Zorin OS 18 (Ubuntu 24.04 Noble / Linux 6.14.0-37-generic x86_64)
• Python: 3.12.3
• Node.js: v24.13.0
• Claude Code CLI: 2.1.34 (system-installed)
• claude-agent-sdk: Tested on 0.1.29, 0.1.30, and 0.1.31 — all reproduce. Pinned back to 0.1.28 as workaround.
• Hardware: Desktop (MS-7C95 motherboard)

The error occurs consistently on every tool call — Read, Glob, Grep, Bash, WebFetch, WebSearch — all produce the same Error in hook callback hook_0: ... Stream closed output. A single planning session with ~25 tool calls generates thousands of lines of noise in the logs.

Confirmed: downgrading to claude-agent-sdk==0.1.28 (bundled CLI v2.1.30) completely eliminates the errors.

[sebyx07]

Cleaned Logs

Here is a cleaned excerpt from a real agent session running with claude-agent-sdk==0.1.31 and permission_mode="bypassPermissions". Every tool call follows the same pattern:

Coding Style Analysis Phase (Opus model)

Using model: opus (claude-opus-4-5-20251101) | Tools: Read, Glob, Grep, Bash, WebFetch, WebSearch

[claude 10:58:16] Using tool: Read → CLAUDE.md
Error in hook callback hook_0: <~50 lines of minified CLI JS source>
error: Stream closed
      at sendRequest (/$bunfs/root/claude:6344:133)
      ...
[claude 10:58:16] Tool completed                              ← succeeds

[claude 10:58:16] Using tool: Glob → *.{eslintrc,prettierrc,editorconfig} in .
Error in hook callback hook_0: <~50 lines of minified CLI JS source>
error: Stream closed
      at sendRequest (/$bunfs/root/claude:6344:133)
      ...
[claude 10:58:16] Tool completed                              ← succeeds

[claude 10:58:16] Using tool: Glob → **/biome.json in .
Error in hook callback hook_0: <~50 lines of minified CLI JS source>
error: Stream closed
      ...
[claude 10:58:16] Tool completed                              ← succeeds

Planning Phase (same session, continues)

[claude 10:58:59] Using tool: Glob → ./tmp/missing-stuff/**/* in .
Error in hook callback hook_0: <~50 lines of minified CLI JS source>
error: Stream closed
      ...
[claude 10:58:59] Tool completed                              ← succeeds

[claude 10:59:00] Using tool: Bash → ls -la ./tmp/missing-stuff/
Error in hook callback hook_0: <~50 lines of minified CLI JS source>
error: Stream closed
      ...
[claude 10:59:00] Tool completed                              ← succeeds

[claude 10:59:04] Using tool: Read → tmp/missing-stuff/0-priority-summary.md
Error in hook callback hook_0: <~50 lines of minified CLI JS source>
error: Stream closed
      ...
[claude 10:59:04] Tool completed                              ← succeeds

Key Observations

1. Every tool type affected: Read, Glob, Grep, Bash — all produce the same error
2. 100% reproduction rate: Not intermittent — happens on literally every tool call
3. Non-fatal: Every tool completes successfully despite the error
4. Minified source dump: The error includes lines 6339-6344 of the bundled Bun CLI source, including the SubagentStop hook handling code:

   hookEvent==="Stop"||hookEvent==="SubagentStop"

5. Stack trace always identical: Always sendRequest → callback → _U1 at the same line numbers
6. Session works fine on v0.1.28: Zero errors, clean output, identical functionality

[sebyx07]

Raw Error Output

This is actual unedited output from a session using query() with permission_mode="bypassPermissions". Every tool call produces this exact block. Here are two consecutive tool calls:

[claude 10:58:16] Using tool: Read → CLAUDE.md
Error in hook callback hook_0: 6339 | Claude Code has been suspended. Run `fg` to bring Claude Code back.
6340 | Note: ctrl + z now suspends Claude Code, ctrl + _ undoes input.
6341 | `)},FL=()=>{l1((SL)=>SL+1)};return uI?.on("suspend",J$),uI?.on("resume",FL),()=>{uI?.off("suspend",J$),uI?.off("resume",FL)}},[uI]);let Cf=lA.useMemo(()=>b5(eA).filter(at),[eA]),kf=lA.useMemo(()=>{if(!ZA)return null;let J$=aL.filter((G9)=>G9.type==="progress"&&G9.data.type==="hook_progress"&&(G9.data.hookEvent==="Stop"||G9.data.hookEvent==="SubagentStop"));if(J$.length===0)return null;let FL=[...new Set(J$.map((G9)=>G9.toolUseID))],SL=FL[FL.length-1];if(!SL)return null;if(aL.some((G9)=>G9.type==="system"&&G9.subtype==="stop_hook_summary"&&G9.toolUseID===SL))return null;let SD=J$.filter((G9)=>G9.toolUseID===SL),BD=SD.length,XI=aL.filter((G9)=>{if(G9.type!=="attachment")return!1;let W1=G9.attachment;return"hookEvent"in W1&&(W1.hookEvent==="Stop"||W1.hookEvent==="SubagentStop")&&"toolUseID"in W1&&W1.toolUseID===SL}).length,R1=SD.find((G9)=>G9.data.statusMessage)?.data.statusMessage;if(R1)return BD===1?`${R1}…`:`${R1}… ${XI}/${BD}`;let yB=SD[0]?.data.hookEvent==="SubagentStop"?"subagent stop":"stop";ret | ... truncated
6342 | `);let A=await NVH(H,$);if(!A.success)throw Error(A.message);if(ND(`${mH.tick} ${A.message}
6343 | `),!A.alreadyUpToDate)c("tengu_plugin_updated_cli",{plugin_id:H,old_version:A.oldVersion||"unknown",new_version:A.newVersion||"unknown"});await Y0(0)}catch(A){lPH(A,`update plugin "${H}"`)}}var QgA=K(()=>{SI();_$();z$();MU();OVH()})
6344 | `)}async sendRequest(H,$,A){let L=HEB.randomUUID(),I={type:"control_request",request_id:L,request:H};if(this.inputClosed)throw Error("Stream closed");if(A?.aborted)throw Error("Request aborted");await this.write(I);let D=()=>{this.write({type:"control_cancel_request",request_id:L});let B=this.pendingRequests.get(L);if(B)B.reject(new oE)};if(A)A.addEventListener("abort",D,{once:!0})

error: Stream closed
      at sendRequest (/$bunfs/root/claude:6344:133)
      at sendRequest (/$bunfs/root/claude:6344:627)
      at <anonymous> (/$bunfs/root/claude:6344:1305)
      at callback (/$bunfs/root/claude:6344:1451)
      at _U1 (/$bunfs/root/claude:3151:1288)
      at _U1 (/$bunfs/root/claude:3151:1525)
      at <anonymous> (/$bunfs/root/claude:3144:1239)
      at A (/$bunfs/root/claude:1240:1491)
      at MK$ (/$bunfs/root/claude:1240:1641)
      at iS (/$bunfs/root/claude:3144:5408)

[claude 10:58:16] Tool completed

[claude 10:58:16] Using tool: Glob → *.{eslintrc,prettierrc,editorconfig,eslintrc.js,eslintrc.json} in .
Error in hook callback hook_0: 6339 | Claude Code has been suspended. Run `fg` to bring Claude Code back.
6340 | Note: ctrl + z now suspends Claude Code, ctrl + _ undoes input.
6341 | <same minified JS source repeated...>
6344 | <same sendRequest / Stream closed pattern...>

error: Stream closed
      at sendRequest (/$bunfs/root/claude:6344:133)
      at sendRequest (/$bunfs/root/claude:6344:627)
      <same stack trace...>

[claude 10:58:16] Tool completed

This repeats identically for every tool call in the session — Read, Glob, Grep, Bash, all affected. A typical agent session has 20-30 tool calls, producing thousands of lines of this error output.