[Bug]: Claude Code SIGABRT shows only "Provider turn failed." while the cause stays in the server log #26082

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

Seen three times on 2026-10-27, plus one deliberate repro, on a Linux container whose cgroup caps tasks at 1000 (pids.max, threads included):

1. Run T3 with the Claude provider in a container with a low task limit (ours: pids.max 1000, 48 visible CPUs).
2. Use a project whose .claude/settings.json has a heavy Stop hook (ours ran a monorepo-wide turbo run lint and turbo run typecheck in parallel, several hundred threads).
3. Start a Claude thread or a Claude delegate_task child and let a turn finish. The Stop hook pushes the cgroup to its limit. Forks fail with EAGAIN, and the Claude Code process aborts.
4. Look at the thread.

The deliberate repro was a trivial claude-haiku-5-5 child task. Its reply arrived, and 14 s later, during the Stop hook, the run failed the same way.

I expect any abnormal Claude Code exit (for example kill -ABRT <claude pid> during a turn) to take the same path, but I haven't verified that shortcut.

Expected behavior

The thread's error item names the cause the server already has: Claude Code process terminated by signal SIGABRT.

When T3's own process spawns fail with EAGAIN at the same moment, a hint that the host ran out of process slots would point users to the real problem. T3 already runs t3-resource-monitor.

Actual behavior

The thread shows only Provider error: Provider turn failed. The cause appears only in ~/.t3/userdata/logs/boot-service.log.

In one thread the run also stayed running for about 6 minutes after the model's end_turn. Claude Code was blocked in the Stop hook, and nothing in the UI said so; then the turn failed with the same bare message.

On the next notification turn, T3 resumed the session into a new Claude Code process. That process aborted within about 3 ms of query.open, again shown only as "Provider turn failed."

Impact

Minor bug or occasional failure

Version or commit

0.1.4-nightly.20261027.3770 (service launcher 0.1.3-nightly.20261027.3754)

Environment

Linux 6.18 container (Railway), cgroup v2 pids.max 1000, Node 24, Claude Code 2.1.311 through the Claude Agent SDK, model claude-opus-5-5 (also claude-haiku-5-5)

Logs or stack traces

[17:19:49.612] WARN (#1142505): orchestration-v2.claude-query-stream-failed
    cause: 'ClaudeAgentSdkQueryRunnerError: Claude Agent SDK query failed.
      at ClaudeAdapterV2.startTurn (…/0.1.4-nightly.20261027.3770/t3:360437:171)
      …
  [cause]: Error: Claude Code process terminated by signal SIGABRT
      at OR.getProcessExitError (…/t3:261049:15)'

[17:19:55.640] WARN (#1510541): thread branch pull request lookup failed
  … [cause]: Error: spawn git EAGAIN

[17:19:57.297] WARN (#1510517): orchestration-v2.claude-mcp-servers-unavailable { servers: [ 't3-code' ] }
[17:19:57.300] WARN (#1510993): orchestration-v2.claude-query-stream-failed
  … [cause]: Error: Claude Code process terminated by signal SIGABRT

The thread item at the same time:

type: error, title: "Provider error", text: "Provider turn failed."

boot-service.log holds 11 terminated by signal SIGABRT failures across several days on this host. Each one showed only "Provider turn failed." in its thread.

Workaround

Grep boot-service.log for claude-query-stream-failed and check /sys/fs/cgroup/pids.events. We cut our Stop hook's thread use so the host stays under its limit.