[BUG] Stream idle timeout / partial response during long tool-use turns on Claude Code Web (Opus 4.7, 1M and non-1M)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Environment

• Platform: Claude Code Web (claude.ai/code)
• Model(s) affected:
• claude-opus-4-7[1m] (1M context) — reproducible
• claude-opus-4-7 (standard context) — also reproducible after switching
• OS: Linux sandbox (provided by the web harness)
• Session type: long-running conversation with multiple tool calls

(Read / Bash / Grep), working in a git repo

• Approx. transcript length when error first appeared: mid-session,

after several dozen tool calls and a few long assistant messages

Summary

During turns where the assistant is about to produce a long text output
(e.g. drafting a ~400-line markdown design doc after a few Read/Bash
tool calls), the stream terminates with:

API Error: Stream idle timeout - partial response received

The error is not triggered by the tool calls themselves — tool results
arrive normally. It consistently fires in the window between the last
tool result and the start (or middle) of the long text reply.

Reproduction

1. Open a Claude Code Web session with Opus 4.7 (1M) on a non-trivial

repo (mine: ~several hundred markdown/py files, custom CLAUDE.md).

1. Hold a long design discussion (tens of turns, many Read/Grep/Bash

calls, several multi-paragraph replies).

1. Ask the assistant to draft a long markdown document (~400+ lines)

into a file, preceded by 1–2 exploratory tool calls.

1. Observe Stream idle timeout - partial response received fire

after the tool calls complete but before / during the long write.

Retry attempts (even after slimming CLAUDE.md and switching from
claude-opus-4-7[1m] to plain claude-opus-4-7) reproduce the same error.

Actual behavior

Stream aborts mid-turn with an idle timeout; partial response is
discarded from the user's perspective and the Write tool call never
executes. The session is usable afterwards, but the same turn cannot
be completed — it fails repeatedly at roughly the same point.

Impact

• Blocks any workflow that involves drafting a long file in a single

turn (design docs, protocol revisions, report packs).

• Forces the user to manually split work into smaller chunks, losing

the model's ability to produce a coherent document in one pass.

• Switching to the non-1M model does not resolve it, so it does not

appear to be strictly a 1M-context issue.

Workarounds tried (none fully effective)

• Slimmed CLAUDE.md to reduce per-turn context overhead — still fails.
• Switched claude-opus-4-7[1m] → claude-opus-4-7 — still fails.
• Starting a fresh session helps temporarily, but the error returns

after the session grows.

Additional context

• The moments when errors occur most frequently are always during the combination of “multi‑turn conversations + about to generate long markdown.”
• The web version cannot call /help, and there is no way to report issues directly from the client, so I am submitting this issue manually.
• Session ID : https://claude.ai/code/session_014AtytC2zxPcAqaoj9LwcGi

Ask

1. Is the idle timeout threshold tunable (e.g. via a server-side

setting or client flag)?

1. Can the stream be kept alive with heartbeats during long

text generation so long Write tool calls don't get cut off?

1. Is there a known interaction between 1M context and long

tail-end text streaming that we should avoid?

What Should Happen?

Expected behavior

The assistant should finish streaming the long reply, or at minimum
fail with a retriable error that preserves the in-progress Write/Edit
tool call.

Error Messages/Logs

API Error: Stream idle timeout - partial response received

Steps to Reproduce

1. Open a Claude Code Web session with Opus 4.7 (1M) on a non-trivial

repo (mine: ~several hundred markdown/py files, custom CLAUDE.md).

1. Hold a long design discussion (tens of turns, many Read/Grep/Bash

calls, several multi-paragraph replies).

1. Ask the assistant to draft a long markdown document (~400+ lines)

into a file, preceded by 1–2 exploratory tool calls.

1. Observe Stream idle timeout - partial response received fire

after the tool calls complete but before / during the long write.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Web (claude.ai/code), encountered on 2026-04-17

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

Additional observation (session 014AtytC2zxPcAqaoj9LwcGi, 2026-04-17):
Stream timeout fired during a ~210-line Edit call. The file write
itself completed successfully (verified via git status post-timeout);
only the client-side stream terminated. This suggests the issue is in
keep-alive / heartbeat between tool-result-accepted and next-assistant-
token, not in the tool execution itself.