| name | harness-tuning |
| description | Tune __HUMAN_NAME__'s shared Codex home-level harness pack safely through the Klimkit repo, not by editing generated files in ~/.codex directly. |
Harness Tuning
Use this skill when HUMAN_NAME asks to change shared Codex behavior, subagents, skills, hooks, model defaults, or other home-level harness files.
Current projected artifact workflow: __KLIMKIT_ARTIFACT_WORKFLOW__; operator folder: __KLIMKIT_OPERATOR_FOLDER__; writable artifact root: __KLIMKIT_ARTIFACT_ROOT__.
This skill runs as one active operator. In team workflow, use other operators' .klimkit/<operator>/ files as readable team context with attribution, and write tuning notes, memories, logs, reflection, task proof, and reports only under the current operator root.
Ground Rules
- Work from
~/klimkit.
- Edit source pack files under
~/klimkit/packs/codex/.
- Do not edit generated projections in
~/.codex/ directly. They are overwritten by kk apply, kk pull, and daemon autosync.
- Keep pack files hand-authored and reviewable. Avoid generated bulk output unless the user explicitly asks for a generated artifact.
- After pack edits, run the relevant pack validation tests and then
kk apply if the current VM should use the change immediately.
- Push committed pack changes to
main when HUMAN_NAME wants the update everywhere. Machines with autosync enabled poll origin/main every 5 seconds by default, fast-forward, apply projections, restart managed services, and send Telegram summaries when configured.
Source Map
packs/codex/AGENTS.md projects to ~/.codex/AGENTS.md.
packs/codex/config.toml projects to ~/.codex/config.toml.
packs/codex/agents/ projects to ~/.codex/agents/.
packs/codex/skills/ projects to ~/.codex/skills/.
packs/codex/hooks/ projects to ~/.codex/hooks/.
Workflow
- Read the existing pack file that owns the behavior.
- For implementation work, create or update the task acceptance checklist before editing source-pack files.
- When integrating external advice, synthesize it into the existing pack instead of pasting template blocks; check
AGENTS.md, subagents, skills, and tests for duplicate or conflicting guidance.
- For UI-impacting workflow changes, require proof reports under the configured writable reports directory (solo
.klimkit/reports/, team __KLIMKIT_REPORTS_PATH__/) with Git-tracked HTML plus screenshot/video evidence, display media as full-width sections, prefer MP4 report videos for Chrome/PWA scrubbing, and include the Tailscale-served report URL in the final handoff when available.
- Make the smallest source-pack edit that solves the request.
- Run
uv run python -m unittest tests.test_codex_pack_validation -q.
- Run broader tests when hooks, config parsing, or Klimkit install/apply behavior changed.
- Run
kk preview for machine-affecting projection changes when useful.
- Run
kk apply to make changes live on the current VM when the user wants the current VM updated now.
- Before final reviewers on non-trivial implementation work, run the Reflection Gate: append a full-timestamped cross-task Reflection Log session to the configured writable reflection file (solo
.klimkit/reflection.md, team __KLIMKIT_REFLECTION_PATH__), using Observations, Derived Pattern, Insight, and Next Probe by default with up to ten named sections when useful; preserve older reflection formats by appending a normalized entry when relevant; reconsider the result; and rerun impacted checks if reflection exposes a gap.
- Commit and push when the change should autosync to other Klimkit machines.
Safety Note
The default pack is tuned for a dedicated VM or external sandbox where yolo-mode Codex is acceptable. Before broadening permissions, check that the VM has minimal secrets, minimal cloud credentials, and only the network/file access needed for the work.