Harbor agent trajectories survive remote sandboxes, but the download silently swallows failures

Two non-obvious facts about harbor (the RL-environment CLI) when you build a pipeline that mines agent transcripts — hint generation, failure analysis, RL trace collection.

1. Remote backends preserve trajectories identically to docker

It is easy to assume --env daytona / --env beam loses in-container logs. It does not. The mechanism is backend-agnostic, in harbor/trial/trial.py::_download_agent_logs:

if self.agent_environment.capabilities.mounted:
    await self.agent_environment.prepare_logs_for_host()   # docker: bind-mounted already
    self._are_agent_logs_downloaded = True
    return
...
await self.agent_environment.download_dir(
    source_dir=self.agent_env_paths.agent_dir.as_posix(),   # /logs/agent
    target_dir=self.paths.agent_dir,                        # <trial_dir>/agent
)

download_dir is an @abstractmethod on BaseEnvironment (harbor/environments/base.py), so every backend implements it. Host layout is <jobs-dir>/<run>/<task-slug>__<shortuuid7>/agent/.

2. The download is best-effort and failure is invisible

The call is wrapped in a bare handler that logs and moves on, then the flag is set unconditionally:

except Exception:
    self.logger.error(f"Failed to download logs to {self.paths.agent_dir}")

self._are_agent_logs_downloaded = True

So a failed or partial log sync does not fail the trial, does not surface in result.json, and does not change the reward. You get a trial that looks completely normal with an empty or truncated agent/ dir. If your pipeline assumes reward-bearing trials carry transcripts, it will quietly produce zero-context results and you will debug the wrong layer.

Do this: treat a missing/empty trial agent/ dir as a first-class case. Count how many trials actually yielded a transcript and assert on it, rather than inferring transcript availability from trial completion. This is the same class of bug as a benchmark stage that degrades to a default value — the numbers still show up, they just no longer mean anything.

3. There is a normalized trajectory.json most tooling ignores

Agents with SUPPORTS_ATIF = True (20+ of them: claude_code, codex, cursor_cli, gemini_cli, copilot_cli, goose, cline, devin, kimi_cli, mini_swe_agent, hermes, grok_build, acp, ...) write a structured ATIF trajectory.json into /logs/agent/ alongside the raw stdout tee (claude-code.txt, codex.txt, ...).

Tooling that globs agent/*.txt silently skips it and ends up regex-scraping JSONL that was already parsed for you. If you need turns, tool calls, or token metrics, read agent/trajectory.json first and fall back to the .txt only for agents where SUPPORTS_ATIF is False.

Also note config.agent.include_logs / exclude_logs switch the sync to download_dir_filtered — an over-narrow include pattern is a second way to end up with a legitimately empty agent/ dir.

Verified against harbor installed via uv tool install harbor (harbor/trial/trial.py, harbor/environments/base.py, harbor/agents/base.py, harbor/agents/installed/*.py).