examples: add swe-agent-harbor-daytona (Harbor sandboxes on Daytona) by Shi-Dong · Pull Request #1919 · radixark/miles

Shi-Dong · 2026-07-28T21:10:11Z

Summary

Adds examples/swe-agent-harbor-daytona, a variant of examples/swe-agent whose
task sandboxes run on Daytona cloud sandboxes instead
of local Docker.

Two things differ from examples/swe-agent:

Daytona sandboxes. The agent-server host needs outbound HTTPS but no
Docker daemon, no local image builds, and no local disk for task images. This
is the practical option when the trainer runs on a GPU node where you cannot
run Docker-in-Docker.
terminus-2 agent. terminus-2 runs as a host process and calls the model
endpoint itself rather than from inside the sandbox. It appends both user
and tool turns, so the session server needs
--tito-allowed-append-roles user tool. Without it every append is rejected,
trajectories collapse to a single turn, and the session server rolls them
back. That flag is the reason this needs its own launcher instead of a README
pointing at examples/swe-agent/run.py.

Everything else — TITO, the session server, GRPO, the reward path — is shared
with examples/swe-agent, so run.py puts that example on PYTHONPATH and
imports its generate.py and swe_agent_function.py rather than keeping a
second copy in sync. Data prep also reuses
examples/swe-agent/download_and_process_data.py, with --agent-name terminus-2.

--rollout-max-response-len and AGENT_MAX_OUTPUT_TOKENS cap a single
turn, not the whole trajectory. Under HARBOR_RESPONSE_LENGTH_POLICY=abort
an over-cap turn ends the trial with none of that turn's tool calls performed,
so the trial scores 0 and dilutes its GRPO group.
Read rollout/raw_reward for the solve rate; rollout/rewards is the
GRPO-centered advantage and sits near zero by construction.
When censusing trial directories by outcome, key on the trial's start time
(the mtime of its config.json), because writing exception.txt bumps the
directory mtime and makes any mtime-sorted listing look like everything is
failing.

Two launcher settings that are hardcoded in examples/swe-agent/run.py are
exposed as flags here, since they are the ones you actually end up tuning:
--rollout-max-response-len and --max-tokens-per-gpu (plus --lr and
--sglang-mem-fraction-static).

--router-external-host defaults to the resolved local IP. sgl-router parses
this into a Rust SocketAddr, so it must be a numeric IP and a hostname fails
to bind. The default is computed in a helper guarded against resolution
failure, so an unresolvable hostname cannot break module import.

Test plan

python -m py_compile examples/swe-agent-harbor-daytona/run.py
bash -n examples/swe-agent-harbor-daytona/launch_agent_server.sh
ruff check examples/swe-agent-harbor-daytona/run.py
Pipeline validated end-to-end by a multi-day GLM-4.7-Flash run on one node
of 8 H200 GPUs against Terminal-Bench 2 tasks, reaching non-zero
rollout/raw_reward with the agent server colocated on the trainer host.
Docs-and-example change; no library code is touched, so no existing test
covers it.

Adds a SWE-agent example whose task sandboxes run on Daytona cloud sandboxes instead of local Docker, so the trainer host needs no Docker daemon and no local task-image builds. The agent is terminus-2, which runs as a host process and appends both user and tool turns, so the session server needs --tito-allowed-append-roles user tool. That flag is the reason this needs its own launcher rather than reusing examples/swe-agent/run.py. generate.py and swe_agent_function.py are unchanged for Daytona-backed runs, so run.py puts the sibling example on PYTHONPATH instead of duplicating them.

gemini-code-assist · 2026-07-28T21:10:13Z

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

- rename the example directory to swe-agent-harbor-daytona - name the harbor branch (harbor-miles-v0.20.0) in launch_agent_server.sh - trim the Daytona disk-quota discussion down to the sizing rule - refer to sibling files by their path from the repo root, not ../swe-agent - drop the troubleshooting row for a missing --tito-allowed-append-roles, which run.py always passes

- name the harbor branch (harbor-miles-v0.20.0) in launch_agent_server.sh - trim the Daytona disk-quota discussion down to the sizing rule - refer to sibling files by their path from the repo root, not ../swe-agent - drop the troubleshooting row for a missing --tito-allowed-append-roles, which run.py always passes - follow the directory rename through paths and default run names

- say up front that the example targets a single node of 8 H200 GPUs, and drop the now-duplicated hardware note from the launch section - reference examples/swe-agent as plain text rather than a relative link - spell out rollout/response_len/max instead of abbreviating it as .../max

nblintao · 2026-08-04T00:05:27Z

@@ -0,0 +1,284 @@
+"""SWE-Agent launcher (GLM-4.7-Flash) for Harbor tasks on Daytona sandboxes.
+
+Same Miles <-> Harbor pipeline as examples/swe-agent, with two differences:


Do you think we should rename swe-agent to something like swe-agent-harbor-docker? or should we should better keep the url slug unchanged?

Good idea. Opened PR 2233 to handle the renaming.

nblintao · 2026-08-04T00:28:11Z

+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --max-concurrent "$MAX_CONCURRENT" \
+    --agent-timeout 5400 \


examples/swe-agent/swe_agent_function.py harcoded to

response = await asyncio.wait_for( _post_agent_server(f"{agent_server_url}/run", request), timeout=3600, # 1 hour max per trial

so the server will still be running after the agent function gives up.

Opened a follow-up PR 2228 to address this comment.

…nt/run.py The launcher only existed because terminus-2 needed --tito-allowed-append-roles. #1818 removed that flag and added a test asserting it is rejected, so the launcher could no longer start a run. Nothing else in it was Daytona-specific: Daytona is selected entirely by the agent server's environment, which the trainer never sees. The four values it exposed as flags all defaulted to what examples/swe-agent/run.py already hardcodes, so reusing that launcher is behaviourally identical. Also corrects --router-external-host: it is substituted into the base URL handed to the agent and only has to resolve from the agent-server host, so a hostname is fine. The bind-time constraint belongs to --miles-host-ip.