Streamline harness install spec names

[xeophon]

Summary

• make version a first-class harness config field via base vf.HarnessConfig.version, alongside fields like max_turns
• add Harness.load_program_config(config) so harnesses can resolve command programs from harness-owned config
• move agent version wiring out of nested program config for OpenCode, MiniSWEAgent, Pi, and Terminus2; configs now use [eval.harness] version = "..."
• keep execution/program knobs on the program config while tests/docs read version-sensitive commands from the resolved harness.program_config
• keep the earlier RLM/composable naming cleanup: shorter shared names like repo_url, ref, tools, exec_timeout, and max_depth

Tests

• uv run --no-sync ruff format
• uv run --no-sync ruff check --fix
• UV_FROZEN=1 uv run --no-sync pre-commit run ty --hook-stage pre-push --all-files
• uv run --no-sync pytest tests/test_v1_harbor_cli.py tests/test_v1_mini_swe_agent.py tests/test_composable_env.py
• uv run --no-sync pytest tests/test_opencode_harbor.py tests/test_imports.py tests/test_v1_config_extension.py

Note

Rename harness install spec fields from package/release to version across all harnesses

• Standardizes version selection across all harnesses (opencode, mini_swe_agent, pi, terminus_2, rlm) by replacing harness-specific field names (package, release, rlm_repo_ref, etc.) with a uniform version field on each harness config.
• Adds a version: str | None field to the base HarnessConfig dataclass and a load_program_config hook on Harness so subclasses can inject version into program.resolve(version=...).
• Renames rlm_-prefixed fields on RLMProgramConfig and rlm_harness (e.g. rlm_tools→tools, rlm_repo_url→repo_url) for consistency.
• Updates all tests and README examples to use the new naming.
• Behavioral Change: config files and code using old field names (package, release, rlm_tools, rlm_repo_url, etc.) will break; no changes to runtime behavior or installed artifacts.

^{Macroscope summarized 1c81dc8.}

---

Note

Medium Risk
Renames public harness and RLM config/TOML fields without backward-compatible aliases, so existing eval configs may silently use defaults; rollout behavior otherwise follows the same install/resolve paths covered by updated tests.

Overview
Standardizes harness configuration around a shared version field for agent installs (OpenCode, Pi, mini-SWE-agent, Terminus 2) and drops per-harness names like release, package, and harbor_package. Harness classes now override load_program_config so that top-level version is passed into each program’s resolve(), and tests/docs read resolved setup via harness.program_config instead of nested program config alone.

RLM and composable RLM helpers rename prefixed knobs to shorter names (repo_url / ref / tools / exec_timeout / max_depth, etc.). HarnessConfig also gains an optional version hook for the base loader pattern.

Breaking: configs or code still using the old field names will not populate the new fields (no aliases in the diff).

^{Reviewed by Cursor Bugbot for commit 1c81dc8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

1 blocking correctness issue found. While mostly mechanical renames, this PR includes structural changes to how version configuration flows through harness classes. Additionally, an unresolved review comment identifies a bug where user-configured program versions could be silently overwritten by defaults.

^{You can customize Macroscope's approvability policy. Learn more.}

Dismissing prior approval to re-evaluate 48ac9e8

Dismissing prior approval to re-evaluate 4027dce

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 2a8e281. Configure here.}