API stream hangs mid-response on complex generation tasks (all models: Opus, Sonnet, Haiku)

Description

Claude Code sessions using Opus 4.5 repeatedly hang mid-stream when attempting complex generation tasks (e.g., writing comprehensive review documents). The API connection remains established but no tokens are received after the initial few.

Environment

Claude Code version: 2.1.14
Model: claude-opus-4-5-20251101
OS: Linux (Debian)
Extended thinking: enabled (high)

Reproduction Steps

Start a session with Opus 4.5 and extended thinking enabled
Build up context (~60-80K tokens) through normal work
Ask the agent to write a comprehensive document (e.g., "write the comprehensive cross-phase consistency review")
Agent responds with thinking like "Let me write the comprehensive review document now"
Stream starts, receives 1-3 output tokens, then hangs indefinitely

Observed Behavior

From session 9b3cc295-c95a-4ca5-b79c-c969ff31b696:

{
  "content": [{"type":"text","text":"Now I have all the information I need. Let me write the comprehensive cross-phase consistency review."}],
  "stop_reason": null,
  "output_tokens": 1
}

TCP connections remain ESTABLISHED to Anthropic servers
Debug logs show "Stream started - received first chunk" then nothing
Process stays in Running state but produces no output
No error messages or retry attempts
Happens repeatedly (261 user interruptions logged due to this pattern)

Debug Log Pattern

[DEBUG] Stream started - received first chunk
[DEBUG] Stream started - received first chunk

Then silence. No further log entries despite process remaining active.

Attempted Fixes

Set MAX_THINKING_TOKENS=63999 (no effect - this is not a token limit issue since stream DOES start)
The request initiates and streams partial text before hanging

Impact

Occurring daily for 3+ days
Affects almost every session involving complex generation tasks
Forces users to interrupt and retry repeatedly
Makes Opus 4.5 unusable for document generation tasks

Request

Please investigate why the API stream stalls mid-response when using Opus 4.5 with extended thinking on larger contexts.

Found 3 possible duplicate issues:

API Streaming Stalls Causing 59-138 Second Delays and Timeouts #18028
[BUG] Opus 4.5 hangs 5-6+ minutes on first prompt in new repo when MCP memory check returns empty #15140
[Bug] Anthropic API: Extended Thinking timeout without user feedback causing CLI hang #13294

This issue will be automatically closed as a duplicate in 3 days.

If your issue is a duplicate, please close it and 👍 the existing issue instead
To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

Update: Also affects Sonnet 4.5

After switching from Opus 4.5 to Sonnet 4.5 in the same session, the identical problem occurs:

Debug log shows model switch:

prev: "Opus 4.5 · Claude Max"
next: "Sonnet 4.5 · Claude Max"

Sonnet 4.5 exhibits the same behavior:

Stream starts, receives first chunk
Gets 3 output tokens
stop_reason: null
Then hangs indefinitely

Additionally seeing API errors with empty error objects:

{
  "type": "system",
  "subtype": "api_error",
  "error": {},
  "retryAttempt": 4,
  "maxRetries": 10
}

The retries also fail with the same pattern.

Conclusion: This is not model-specific. Both Opus 4.5 and Sonnet 4.5 hang mid-stream on this session. The issue may be related to:

Session context corruption
Extended thinking configuration persisting across model switch
API-side issue affecting both models

Updating title to reflect this affects multiple models.

Update 2: Also affects Haiku

The issue also reproduces with Haiku (claude-haiku-4-5-20251001). Same behavior:

Stream starts
Few tokens received
Hangs mid-response

This affects all three model families:

Opus 4.5
Sonnet 4.5
Haiku 4.5

Not model-specific - appears to be an API-level or session-level issue.

Additional Debug Information

Error Pattern

The actual error captured in debug logs:

[ERROR] Error in non-streaming fallback: Request was aborted.
[ERROR] Error: Error: Request was aborted.
    at D (/$bunfs/root/claude:1028:26168)
    at A4H (/$bunfs/root/claude:1028:26102)

Retry Escalation

The session shows exponential backoff retries, all failing:

20:20 - retry 1 (529ms delay)
20:25 - retry 2 (1037ms)
20:30 - retry 3 (2039ms)
20:35 - retry 4 (4557ms)
20:40 - retry 5 (9294ms)
20:45 - retry 6 (16559ms)

Critical Issue: Empty Error Objects

All API errors are logged with "error": {} - the actual error details are not being captured:

{
  "type": "system",
  "subtype": "api_error", 
  "error": {},
  "retryAttempt": 6,
  "maxRetries": 10
}

This makes debugging nearly impossible. The error object should contain the actual API response/error code.

Summary

Stream starts, receives 1-3 tokens
"Request was aborted" error in non-streaming fallback
Retries with exponential backoff (up to 10 retries)
Each retry fails with same pattern
Error details not captured (error: {})
Affects all models (Opus, Sonnet, Haiku)
Session becomes permanently stuck in retry loop

Update: Occurs with Extended Thinking OFF

The issue also reproduces with extended thinking disabled. This rules out the MAX_THINKING_TOKENS off-by-one bug as the cause.

Confirmed NOT related to:

Extended thinking settings
MAX_THINKING_TOKENS limit
Specific model (affects Opus, Sonnet, Haiku)

The stream abortion happens regardless of thinking mode configuration.

Additional Diagnostic Data (macOS)

Reproducing the "stream hang at 0 tokens" issue on macOS with Claude Code 2.1.17.

Environment

Claude Code version: 2.1.17
Model: claude-opus-4-5-20251101
OS: macOS 26.2 (Darwin 25.2.0)
Date: 2026-02-01

Symptoms

Same as OP: "Julienning" spinner with ↓ 0 tokens for 10+ minutes. User must ESC to interrupt.

Key Finding: TCP Connection Degradation

Unlike issues where the stream simply stops, this case shows severe TCP-level packet loss on an established connection:

tcp4 172.17.12.178:52525<->160.79.104.10:443
State:          ESTABLISHED
bytes_in:       ~154KB (partial response)
bytes_out:      ~1.1MB (request sent)
re-tx:          25,983  ← CRITICAL: 25k+ retransmissions
rx_dupe:        1,187
rx_ooo:         271

For comparison, healthy connections to the same API endpoint in the same session show 0 retransmissions.

Network Path is Clean

Traceroute shows 8 hops with normal latency (~22ms), 0% packet loss on ping:

1  172.17.12.1  7ms
...
8  160.79.104.10  22ms

Process State

sample shows main thread blocked in kevent64 (78% of samples) - correctly waiting for network events that aren't arriving due to the degraded TCP state.

Theory

The high retransmission count suggests the TCP connection entered a degraded state mid-stream where packets are being lost or corrupted. The kernel keeps retrying, but the connection is effectively stalled. This would explain why:

Stream starts successfully (first chunk arrives)
Then hangs (subsequent data doesn't arrive/ack properly)
No error is raised (TCP stack thinks connection is still valid)
Eventually times out after TCP keepalive exhaustion

Possible Causes

Middlebox (firewall/proxy) interfering with long-lived HTTPS connections
API server-side issue with response streaming
TLS session corruption mid-stream
Connection idle timeout from network equipment

Reproduction Note

Multiple Claude Code sessions on same machine showed similar degradation pattern today (PID 82918: 25k retx, PID 34314: 10k retx). Fresh connections work fine.

Diagnostics collected while leaving the hung process running for analysis

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.

Navigation Menu

Repository navigation

API stream hangs mid-response on complex generation tasks (all models: Opus, Sonnet, Haiku) #19883

Description

Environment

Reproduction Steps

Observed Behavior

Debug Log Pattern

Attempted Fixes

Impact

Request

Update: Also affects Sonnet 4.5

Update 2: Also affects Haiku

Additional Debug Information

Error Pattern

Retry Escalation

Critical Issue: Empty Error Objects

Summary

Update: Occurs with Extended Thinking OFF

Additional Diagnostic Data (macOS)

Environment

Symptoms

Key Finding: TCP Connection Degradation

Network Path is Clean

Process State

Theory

Possible Causes

Reproduction Note

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

API stream hangs mid-response on complex generation tasks (all models: Opus, Sonnet, Haiku) #19883

Description

Description

Environment

Reproduction Steps

Observed Behavior

Debug Log Pattern

Attempted Fixes

Impact

Request

Activity

github-actions commented on Jan 21, 2026

josh-argyle commented on Jan 21, 2026

Update: Also affects Sonnet 4.5

josh-argyle commented on Jan 21, 2026

Update 2: Also affects Haiku

josh-argyle commented on Jan 21, 2026

Additional Debug Information

Error Pattern

Retry Escalation

Critical Issue: Empty Error Objects

Summary

josh-argyle commented on Jan 21, 2026

Update: Occurs with Extended Thinking OFF

spm1001 commented on Feb 1, 2026

Additional Diagnostic Data (macOS)

Environment

Symptoms

Key Finding: TCP Connection Degradation

Network Path is Clean

Process State

Theory

Possible Causes

Reproduction Note

github-actions commented on Mar 2, 2026

github-actions commented on Mar 10, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions