[BUG] Cowork: all tasks fail with retry loop then ECONNRESET — fails across networks incl. phone hotspot, survived clean reinstall, worked hours earlier same day (macOS 15.7.3, M3)

I have also been encountering this issue for multiple days now. It's getting beyond frustrating. Have contacted support with zero response.

<img width="633" height="422" alt="Image" src="https://github.com/user-attachments/assets/878371da-a423-4aac-87fb-8d534c62a75d" />

Hello ,
I have been facing this for several weeks now and today . I switched to Hotspot and 2G WIFI and discovered that I dont get the Server is busy error and the above error connection message anymore . I have switched back and forth between my regular 5G and the aforementioned networks and I am able to replicate this error . Strangely , this is only a co-work problem and not a chat/code problem . I have also switched between Mac and Windows ( currently using Mac Pro M1 ) .

<p style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Adding another reproduction with evidence the OP's report doesn't have: host logs showing an<span class="Apple-converted-space"> </span><strong>API-sent<span class="Apple-converted-space"> </span><code>server_error</code>event preceding the retry loop</strong>, plus two independent HTTP 529 errors from Anthropic's auth service in the same period, with request IDs Anthropic can trace server-side.</p><h2 style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Environment</h2><ul style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;"><li>macOS (arm64, Apple Silicon)</li><li>Claude Desktop 1.18286.0 (Jul 5 failures) and 1.19367.0 (Jul 8 failure)</li><li>Claude Code SDK in VM: 2.1.197</li><li>Model: claude-fable-5</li><li>No VPN, no third-party firewall/network filters</li><li>Network: residential (Central California)</li></ul><h2 style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Failure signature</h2><p style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Three failed Cowork turns across two dates, all identical:</p>
Date/time (local) | seconds_to_outcome | had_first_response | transcript_size_bytes
-- | -- | -- | --
2026-07-05 02:36:05 | 179 | true | 75,252
2026-07-05 02:40:02 (resume of same task) | 192 | true | 80,683
2026-07-08 01:28:37 | 185 | true | 84,103

<p style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Sequence in every case: task starts and streams normally → API returns<span class="Apple-converted-space"> </span><code>server_error</code><span class="Apple-converted-space"> </span>mid-stream → UI shows "Server is busy. Retrying in Ns (attempt N of 10)" → all retries fail → final<span class="Apple-converted-space"> </span><code>ECONNRESET</code>. The consistent ~180–190s outcome time is the retry-exhaustion window, not a fixed network timeout.</p><pre style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;"><code>2026-07-05 02:36:05 [info] [APIError] Intermediate SDK error "server_error" for session local_2489793f-... — continuing to read stream
2026-07-05 02:36:05 [info] [Result] Turn failed for session local_2489793f-...: subtype=success, is_error=true
2026-07-05 02:36:05 [info] [CycleHealth] Unhealthy cycle: {
session_id: 'local_2489793f-67b7-4022-9460-f8c3fef2e490',
variant_key: 'marigold',
host_loop_mode: true,
model: 'claude-fable-5',
cycle_health: 'unhealthy',
had_first_response: true,
seconds_to_outcome: 179,
is_resume: false,
unhealthy_reason: 'api_error',
error_message: 'API Error: Unable to connect to API (ECONNRESET)',
transcript_size_bytes: 75252
}
</code></pre><pre style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;"><code>2026-07-08 01:28:37 [info] [APIError] Intermediate SDK error "server_error" for session local_66be0866-... — continuing to read stream
2026-07-08 01:28:37 [info] [CycleHealth] Unhealthy cycle: {
session_id: 'local_66be0866-136d-46fa-b0cd-2bdcf916a7e5',
model: 'claude-fable-5',
had_first_response: true,
seconds_to_outcome: 185,
unhealthy_reason: 'api_error',
error_message: 'API Error: Unable to connect to API (ECONNRESET)',
transcript_size_bytes: 84103
}
</code></pre><h2 style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">VM/networking healthy at every failure</h2><p style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">coworkd.log shows a clean bootstrap on every boot in this window — static IP, MTU 1400, ephemeral CA generated and installed, MITM proxy started, and reachability to https://api.anthropic.com passing every time:</p><pre style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;"><code>[coworkd] lowered MTU from 1500 to 1400 on enp0s1
[proxy] MITM proxy started on /var/run/mitm-proxy.sock
[coworkd] checking reachability https://api.anthropic.com/ (timeout=30s)
[coworkd] reachability check passed (status=404)
</code></pre><h2 style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Corroborating server-side errors (traceable request IDs)</h2><p style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Independently of Cowork tasks, the host app logged two HTTP 529 errors from Anthropic in the same period:</p><pre style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;"><code>2026-06-30 23:04:22 [error] [sessions-bridge] Poll error, backing off: Poll: Failed with status 529: Authentication service is temporarily unavailable. Retry the request. [request-id: req_011CcaqFRZgMwk6wbos2M72e]
2026-07-07 10:22:40 [error] [sessions-bridge] Poll error, backing off: Poll: Failed with status 529: Authentication service is temporarily unavailable. Retry the request. [request-id: req_011Cco5pxPvZnZSvnXWg6jaU]
</code></pre><h2 style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Why this matters for triage</h2><p style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">The failure is episodic, not persistent: sessions run minutes after a failure (2026-07-05 02:41 and 02:54) completed without error. Combined with the API-sent<span class="Apple-converted-space"> </span><code>server_error</code><span class="Apple-converted-space"> </span>initiating each retry loop and the 529s from the auth service, this points at the API/edge side rather than client VM networking — the ECONNRESET appears to be the terminal symptom of retries against an already-unhealthy upstream, not the root cause. The two request IDs above should be traceable in Anthropic's logs.</p><p style="caret-color: rgb(255, 255, 255); color: rgb(255, 255, 255); font-style: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration-line: none; text-decoration-thickness: auto; text-decoration-style: solid;">Happy to provide full main.log / coworkd.log excerpts on request.</p>

Follow-up to my comment above: the failure has now survived a complete clean reinstall (app removed, ~/Library/Application Support/Claude deleted, VM bundle re-downloaded fresh) and reproduced on 2026-07-15 with the same signature. New evidence below, including the mid-stream server_error that precedes the ECONNRESET.

The key finding: the ECONNRESET is the terminal symptom, not the root cause. The triggering event is a server-side server_error received mid-stream, and the failures cluster in a consistent ~180–210 second window — the OP's log shows seconds_to_outcome: 186; mine shows 208.

Environment

• macOS, Mac Studio (Apple Silicon, 64 GB)
• Claude Desktop, fresh install 2026-07-15
• Model: claude-fable-5
• Failure timestamp: 2026-07-15 15:51:40 (Pacific)

Identifiers for backend traceability

• session_id: local_0b03b2dd-f6f7-4fde-954a-be841ef51601
• cli_session_id: 53096442-8f70-4f78-8a28-21f4c7df25eb
• vm_instance_id: 22841919-1903-493b-ad97-b27b46121271
• user_message_uuid: 4d85acb2-5a84-404b-94f1-61eff05d8ef5

Log sequence at failure (main.log, 2026-07-15)

15:51:36 [remote-tools-device] socket closed: 1006
15:51:36 [remote-tools-device] reconnect #1 in 1000ms
15:51:37 [remote-tools-device] socket closed: 1006
15:51:37 [remote-tools-device] reconnect #2 in 1500ms
15:51:40 [APIError] Intermediate SDK error "server_error" for session
         local_0b03b2dd-... — continuing to read stream
15:51:40 [Result] Turn failed: subtype=success, is_error=true
15:51:40 [CycleHealth] Unhealthy cycle: {
  cycle_health: 'unhealthy',
  unhealthy_reason: 'api_error',
  error_message: 'API Error: Unable to connect to API (ECONNRESET)',
  had_first_response: true,
  seconds_to_outcome: 208,
  is_resume: true,
  transcript_size_bytes: 82283,
  host_loop_mode: true,
  permission_mode: 'auto'
}

Observations

1. Not a connection-establishment failure. had_first_response: true — the session connected, streamed for 208 seconds, then died. Matches the OP's 186s. Across my July 5–15 failures the outcome consistently lands ~180–210s.
2. The API sent a server_error mid-stream 4 seconds before the turn failed; the client kept reading and the connection was then reset. The UI reports only the ECONNRESET, which is likely why this bug is being filed under multiple different titles.
3. A second, independent connection failed in the same window. The remote-tools-device websocket to bridge.claudeusercontent.com dropped twice with code 1006 (abnormal closure) at 15:51:36–37, seconds before the API stream died. Two separate connections to Anthropic infrastructure failing simultaneously is consistent with an upstream/edge event, not a local network fault.
4. Local causes exhausted (corroborating the OP's ruling-out):

• Host path clean: curl -v https://api.anthropic.com/v1/messages → TLS 1.3 OK, Cloudflare responds (expected 405)
• systemextensionsctl list → 0 network system extensions; macOS firewall off; no VPN/proxy/DNS filtering active during repro
• vmnet unified log during the repro window: empty (VM NAT healthy)
• Identical failure before and after clean reinstall

1. Session context: large MCP tool surface (~10 servers, one exposing 116 tools) and is_resume: true with an 82 KB transcript — i.e., long streaming turns that routinely extend past the ~3-minute mark where these failures cluster. May be relevant to why some users hit this constantly and others never do.

Happy to provide full main.log/coworkd.log excerpts or run an instrumented repro.

Has any resolution been offered for this? I am experiencing theexact same symptoms.