Bump harnesses verifiers minimum

[xeophon]

Replacement for closed #1545.

Stacked on #1544 by branch ancestry, while still targeting main: harnesses-verifiers-dev158 is one commit on top of harnesses-0.1.2-main (50bf7c27).

This bumps the standalone harnesses package dependency from verifiers>=0.1.15.dev11 to verifiers>=0.1.15.dev158, so harnesses==0.1.2 only installs with a verifiers build that includes the Harness.load_program_config hook used by the versioned harness config changes.

Addresses #1544 (comment).

Verification:

• UV_FROZEN=1 uv build --out-dir /tmp/harnesses-verifiers-dev158-restacked packages/harnesses
• Confirmed wheel metadata contains Requires-Dist: verifiers>=0.1.15.dev158

---

Note

Medium Risk
Breaking config renames (package/release/rlm_* fields) affect eval TOML and Python callers; dependency floor ties releases to verifiers dev158+.

Overview
Raises harnesses’ minimum verifiers dependency to >=0.1.15.dev158 and bumps environment packages to harnesses>=0.1.2, so harnesses==0.1.2 only installs against a verifiers build that includes Harness.load_program_config.

Harness.__init__ now resolves programs via load_program_config instead of calling config.program.resolve() directly; command harnesses (OpenCode, Pi, MiniSWEAgent, Terminus2) override it to pass a top-level version into program.resolve(version=...). Per-program package / release fields are removed in favor of HarnessConfig.version; docs show [eval.harness] with id + version.

RLMProgramConfig and rlm_swe_v1 rename rlm_* knobs (rlm_tools→tools, rlm_repo_url→repo_url, etc.). Tests assert harness.program_config for resolved install scripts.

^{Reviewed by Cursor Bugbot for commit 2dcd264. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Move version configuration to harness-level in verifiers and harnesses packages

• Moves version/release fields from program config dataclasses (MiniSWEAgentProgramConfig, OpenCodeProgramConfig, PiProgramConfig, Terminus2ProgramConfig) to their parent harness config dataclasses as a top-level version field.
• Adds load_program_config as an extension point on the base Harness class so subclasses can inject the harness-level version during program resolution.
• Renames rlm_*-prefixed fields in RLMProgramConfig and RlmSweProgramConfig (e.g. rlm_tools → tools, rlm_exec_timeout → exec_timeout) and updates utility functions in rlm_utils.py to match.
• Bumps the verifiers minimum from >=0.1.15.dev11 to >=0.1.15.dev158 and harnesses minimum from >=0.1.1 to >=0.1.2 across environment packages.
• Risk: Breaking config change — existing configs using program.release, program.package, program.harbor_package, or rlm_tools/rlm_exec_timeout field names will no longer be valid.

^{Macroscope summarized 2dcd264.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR restructures the configuration API across multiple harness implementations, moving version specification from ProgramConfig to HarnessConfig and introducing a new load_program_config method pattern. The cross-cutting nature of these API changes to shared infrastructure warrants human review.

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