Add examples/swe-agent: GLM-4.7-Flash agentic training with Harbor

[Shi-Dong]

Adds examples/swe-agent: a GLM-4.7-Flash agentic training example that runs synchronous GRPO in Miles while a Harbor agent server creates the task sandboxes, runs the agents, and returns verifier rewards.

This PR is purely additive — 7 new files, no deletions and no edits to existing examples.

Why

Miles currently has no Harbor agentic training example outside examples/experimental/. Every Harbor-based training example lives under examples/experimental/: swe-agent-v2, swe-agent-v2-amd, and openenv. The only Harbor reference outside experimental/ is examples/eval/terminal_bench_via_agent_server, which is evaluation-only and does no training.

That matters because this is the reference multi-turn agentic setup on the session-server TITO path — docs/user-guide/agentic-chat-template.md and docs/user-guide/rollout-endpoints.md both send readers here as the worked example of a custom generate function, a custom agent function, and the abort hook. Users following those docs land in experimental/, in a directory that also contains an unrelated GSM8K reasoning launcher and a launcher that depends on a private repository they cannot clone.

Naming

The directory is swe-agent, not swe-agent-v2. The v2 suffix only made sense inside experimental/, where it separated this Harbor-based generation from the older Nemo-Gym-based experimental/swe-agent. Neither of those is being promoted, so a version suffix at the top level would point readers at a v1 that does not exist there.

What is in the example

File	Purpose
run.py	Synchronous GLM-4.7-Flash launcher (the validated recipe).
run-glm47-flash-agentic-async.py	Disaggregated fully asynchronous launcher.
swe_agent_function.py	Sends each rollout to the Harbor agent server; implements the abort hook.
generate.py	Builds rewards, metrics, and training samples.
download_and_process_data.py	Converts supported datasets to Miles JSONL.
README.md	End-to-end recipe: agent server, data prep, launch, verification.

It points at the public harbor-framework/harbor branch harbor-miles-v0.20.0, so the recipe is runnable from a clean checkout. Agent-server hostnames, checkpoint paths, and the W&B project name are placeholders rather than the values the validation run happened to use.

Both launchers route rollouts through sgl-router. The deprecated --use-miles-router flag that the experimental copy still passes has been dropped from both, matching the configuration the validation run below actually used (use_miles_router=False, router_backend=sglang).

Validation

A GLM-4.7-Flash Terminal-Bench 2 run on one node of 8xH200 has been training continuously on these files for several days and is past 124 synchronous GRPO steps, with checkpoints saved on schedule and TITO session tracking healthy throughout. The README documents that run's actual shape: 4 prompts x 8 samples = 32 trajectories per step, --global-batch-size 32, --max-seq-len 65536, --num-rollout 200, --save-interval 20. run.py defaults match.

That run surfaced two operational lessons the README passes on in general terms. Synchronous rollout waits for the slowest trajectory in the batch, so a step that draws an unusually slow task can take several times the typical 10 minutes. And W&B can silently stop accepting some metric rows while others keep arriving, which looks exactly like a frozen reward curve — so "Verify progress" says to confirm a suspected stall against the on-disk train_data/<step> and rollout_data/<step>.pt dumps rather than a dashboard.

The README also explains how to size the agent server's --max-concurrent (one sandbox per trajectory in a step) and why --agent-timeout must be generous.

Follow-up

Deleting the now-duplicated examples/experimental/swe-agent-v2 and repointing the two docs pages at examples/swe-agent will come as a separate PR, so this one stays reviewable as an addition.

Test plan

• Long-running GLM-4.7-Flash Terminal-Bench 2 training run on these files (124+ GRPO steps, one node, 8xH200)
• python -m py_compile on all launchers
• ruff check examples/swe-agent/
• git diff --check
• Net diff against main confirmed additive-only (no deletions, no modifications to existing files other than the examples/README.md index entry)

[gemini-code-assist]

Code Review

This pull request promotes the swe-agent-v2 example from the experimental directory to the main examples directory, updating all documentation and relative paths. Additionally, it refactors the launcher script run.py to introduce configurable parameters for rollout limits, save intervals, and trace directories, while updating default settings such as increasing max_seq_len to 65536. Feedback is provided regarding the removal of the --use-miles-router flag in run.py, which could unintentionally bypass the Miles Router and disrupt Token-In-Token-Out (TITO) multi-turn agentic rollouts.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[guapisolo]

Approve to unblock