BenchFlow pipe-closed on Daytona DinD compose tasks — agent install/exec fails silently

[xdotli]

Summary

When running tasks that have docker-compose.yaml on Daytona, BenchFlow fails with Process closed stdout (rc=None) while Harbor succeeds with the same environment. The agent is never installed and never starts.

Reproduction

Same task (gh-repo-analytics), same model (claude-opus-4-7), same Daytona:

Method	Result	Tool Calls	Error
BenchFlow + Docker (host)	✅ reward=0.0	8	None
Harbor + Daytona DinD	⚠️ reward=0.0	N/A	NonZeroAgentExitCodeError (auth issue, but compose worked)
BenchFlow + Daytona DinD	❌	0	Process closed stdout (rc=None)

Harbor proves the DinD compose infrastructure works on Daytona. BenchFlow's ACP path fails at the agent install/exec step.

Root cause analysis

BenchFlow's DaytonaProcess.from_harbor_env() detects DinD mode and constructs docker compose exec commands to run inside the DinD container. However:

1. install_agent() calls env.exec() which goes through Harbor's DaytonaEnvironment._sandbox_exec() → _strategy.exec() → _DaytonaDinD.exec() → _compose_exec(["exec", ...])
2. The compose exec command runs inside the DinD VM via SSH
3. The agent install produces empty stdout — the install command output is lost
4. After install, connect_acp() creates a DaytonaProcess and tries to start the agent via docker compose exec ... claude-agent-acp
5. The agent process immediately closes stdout → Process closed stdout (rc=None)

Likely causes:

• The compose env vars (_compose_env_vars()) may not include the subscription auth credentials needed by the agent
• The SSH → DinD → compose exec chain may lose the --env-file or env var injection
• Node.js may not be installed inside the compose container (the DinD VM is docker:28.3.3-dind Alpine, but the compose main container is built from the task's Dockerfile)

Evidence

BenchFlow DinD result:

• install-stdout.txt: empty
• claude_agent_acp.txt: empty
• timing.json: environment_setup: 42.7s (compose up succeeded)
• config.json: _BENCHFLOW_SUBSCRIPTION_AUTH: 1 (OAuth detected)

Environment

• BenchFlow v0.3.1 (dev-0.3 branch)
• Harbor (pip installed, used by BenchFlow)
• Daytona cloud
• Tasks affected: all 5 with docker-compose.yaml in SkillsBench (gh-repo-analytics, pedestrian-traffic-counting, pg-essay-to-audiobook, scheduling-email-assistant, react-performance-debugging)

Expected behavior

BenchFlow should install and run claude-agent-acp inside the DinD compose main container, same as Harbor installs and runs claude CLI.