Support runtime skill sources through harness skills - #2556
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![macroscopeapp[bot]](https://avatars.githubusercontent.com/in/900172?s=60&v=4){kind=small}
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 541356a373
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
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 541356a. Configure here.
ApprovabilityVerdict: Not approved Macroscope's review found this PR not approvable — This PR adds runtime- and task-provided skills and changes installation, isolation, cleanup, and caching behavior across eight production harnesses and Harbor task loading. The cross-cutting runtime impact and unresolved concerns around validation, reuse, cleanup, and permissions require human review. Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b74c5ae728
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
1 participant
Allow tasks and harnesses to use skills already present in the runtime alongside existing local skill uploads. Harbor's
environment.skills_dirreaches native harness discovery through the shared skills installer.HarnessConfig.skillsandTaskData.skillsaccept local skill folders and runtime directories containing skill folders:Task sources install before harness sources, so configured harness files take precedence when paths overlap. Missing or invalid source directories fail the run.
Each run installs skills in its own native configuration directory or explicit discovery path. RLM installs editable skill packages into a separate uv tool environment for the run and removes that environment during cleanup. Skill-bearing runs use separate harness instances to keep installation state local to the run.
Harbor translates the solver's skills declaration into
TaskData.skills, clears it from verifier task data, and accepts the redundant agent-skills field in deterministic verifier environments.Note
Medium Risk
Touches skill installation across many harnesses and rollout harness cloning; behavior changes for shared vs per-run skill paths (OpenClaw, RLM) could affect concurrent or cached runtime reuse.
Overview
Adds
RuntimeSkillsand a sharedSkillSourcetype so harness and task config can reference host skill folders or directories already inside the agent runtime, with Harbor wiringenvironment.skills_dirintoTaskData.skills.Agent._rollout_paramsmerges task skills before harness skills and clones the harness per run when any skills are configured, so installation state (notably RLM’s uv environment) stays run-local.install_skillscopies runtime roots in place and still uploads host folders;skill_destinationkeeps Pi/Prime Agent--skillpaths aligned with what was installed.Skill-bearing ACP harnesses defer installation from shared
setuppaths to trace-scoped discovery dirs at session prep (Claude, Codex, Kimi, OpenClaw, Pi, Pool, Prime Agent, RLM). OpenClaw drops its shared staged-skills cache; RLM uses a run-specific install tree when skills are set and cleans it up. Several harnesses addcleanupfor per-trace state dirs. Docs describe TOML{runtime = "..."}entries andTaskData.skillsprecedence.Reviewed by Cursor Bugbot for commit 8450325. Bugbot is set up for automated code reviews on this repo. Configure here.
Note
Add
RuntimeSkillssupport and move skill installation to per-run harness setupRuntimeSkillsPydantic model to represent skill directories already inside the agent runtime, and changesHarnessConfig.skillsto accept both hostPathandRuntimeSkillssourcesskillsfield toTaskData;Agent._rollout_paramscombines task-level and harness-level skill sources in task-then-harness order before validation and executionskill_destinationhelper resolves runtime sources to the destination root and host sources to a named child directory, so all harnesses install skills consistentlysetupphase into per-traceprepare_acporbuild_env, giving each run its own skills directoryparse_tasknow maps the environmentskills_dirto aRuntimeSkillssource, andparse_verifier_environmentno longer rejectsskills_dirsetupno longer installs skills; all skill installation now happens per-run in trace-specific directories and is cleaned up after each run. Reviewers should checkHarness.install_skillsin harness.py and each harness'sprepare_acp/build_envto verify destination resolution matches expectationsMacroscope summarized 8450325.