rollout: decouple oversampling-abort teardown into a pluggable agent hook

[Shi-Dong]

Problem

With oversampling on, once the trainer has collected enough samples, the rollout aborts the in-flight SGLang generations via /abort_request {"abort_all": true}. But the external agent loops driving the agentic rollouts are never told.

Each aborted generation just looks like one finished LLM call to the agent, so the agent immediately issues the next completion request back to SGLang. The loops keep hammering SGLang and holding their sandbox containers until they independently hit their own max_seq_len / trial timeout — wasting inference during the drain and pinning containers.

Why not a hard-coded flush in core

Miles' core rollout is agent-host-agnostic: agentic_tool_call.generate delegates all agent-host knowledge to a plugin loaded from --custom-agent-function-path. Baking a Harbor-specific /flush call into the core rollout would wrongly assume every deployment uses a Harbor agent server that speaks /flush, which isn't always true.

Fix (pluggable hook, both sync and async paths)

• Core stays agnostic. New miles.utils.misc.call_agent_abort_hook(args) looks for a sibling abort callable in the same module as the configured --custom-agent-function-path and awaits it if present. No hook → drain as before. Zero Harbor strings, no new core args. Non-agentic and non-Harbor runs are unaffected.
• Wired into BOTH oversampling-abort implementations:
  • miles/rollout/sglang_rollout.py::abort — the legacy/sync path (default when MILES_EXPERIMENTAL_ROLLOUT_REFACTOR is unset).
  • miles/rollout/inference_rollout/inference_rollout_train.py::abort — the experimental-refactor path used by async training (train_async.py + MILES_EXPERIMENTAL_ROLLOUT_REFACTOR=1 → InferenceRolloutFn). This is the path Shi's agentic async runs actually take, so covering it is the point.
• Harbor teardown lives in the plugin. The Harbor /flush logic is in examples/experimental/swe-agent-v2/swe_agent_function.py as its abort(args) hook — next to the AGENT_SERVER_URL + /run code it already owns. It flushes the agent server's in-flight /run trials for this run's session_server_instance_id (sending HARBOR_ADMIN_SECRET as a bearer token when present). No-op unless the URL + instance id are available.

No Harbor server change is needed — /flush already exists in harbor-miles-v0.13.1 and returns HTTP 200 on cancel.

Test plan

• Sync agentic run (legacy path): on oversampling abort, plugin logs Flushed agent server ... and Harbor /run tasks terminate instead of continuing to call SGLang.
• Async agentic run (train_async.py + refactor): same behavior via inference_rollout_train.abort.
• Agentic run whose plugin defines no abort: call_agent_abort_hook is a silent no-op; drain unchanged.
• Non-agentic (rule-based RM) run: custom_agent_function_path is None → hook lookup short-circuits; abort behavior unchanged.

[gemini-code-assist]

Code Review

This pull request introduces the ability to flush in-flight Harbor agent-server trials when aborting SGLang rollouts. It adds a new --agent-server-url command-line argument and a corresponding flush_agent_server helper function that sends a POST request to the agent server's /flush endpoint. Feedback suggests stripping leading and trailing whitespaces from the agent server URL and the admin secret to prevent potential issues with malformed URLs or invalid authorization headers.

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.

[gemini-code-assist]

Environment variables and command-line arguments (especially URLs and secrets/tokens) are highly prone to containing accidental leading or trailing whitespaces or newlines (e.g., when loaded from .env files or Kubernetes secrets).

Why this matters:

• URLs: Leading/trailing spaces can cause httpx to raise malformed URL or connection errors.
• Secrets: Trailing newlines in HARBOR_ADMIN_SECRET will result in an invalid Authorization header, causing the Harbor agent server to reject the request with a 400 or 401 error.

Stripping these values before use ensures robust and reliable API calls.

Suggested change

	agent_server_url = getattr(args, "agent_server_url", None) or os.environ.get("AGENT_SERVER_URL")
	instance_id = getattr(args, "session_server_instance_id", None)
	if not agent_server_url or not instance_id:
	return
	headers = None
	admin_secret = os.environ.get("HARBOR_ADMIN_SECRET")
	if admin_secret:
	headers = {"Authorization": f"Bearer {admin_secret}"}
	agent_server_url = getattr(args, "agent_server_url", None) or os.environ.get("AGENT_SERVER_URL")
	instance_id = getattr(args, "session_server_instance_id", None)
	if not agent_server_url or not instance_id:
	return
	agent_server_url = agent_server_url.strip()
	headers = None
	admin_secret = os.environ.get("HARBOR_ADMIN_SECRET")
	if admin_secret:
	headers = {"Authorization": f"Bearer {admin_secret.strip()}"}