Skip to content

About

OMP extension to improve workflow and adjust external skills

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

skill-observer

An OMP extension for adjusting upstream skills without editing their installed files.

Your adjustments live in separate amendments that survive skill updates. The extension delivers them when you invoke /skill:<name> or the agent reads the skill's SKILL.md.

Run /observe to review session corrections and proposed fixes. After an upstream skill changes, the next /observe checks whether its amendments still apply. It can also propose local skill improvements and project checks.

Status: vibe-coded experiment.

Requirements and first run

Requires OMP with a configured model and Git on PATH. Install from the repository, then start OMP in your project:

omp install https://github.com/Razunter/skill-observer.git
omp

OMP installs and enables the extension for your user; no manual clone or configuration entry is needed. The extension is not published on npm and needs no runtime npm packages or build step. See setup for checkout loading, OMP's /extensions controls, and state locations.

Use it

After working with a skill and correcting the agent, enter:

/observe

The model reviews session evidence and previews proposed changes with Apply, Later, or Decline choices. With nothing to review in a fresh session, the command reports that and starts no model turn.

You can add feedback or a focus:

/observe Focus on the upstream skills I corrected in this session.

A note starts a review without other signals, but the pass can finish without proposals if evidence does not support a change. Observation runs only on request, never automatically at session exit.

To see what the extension knows about the current branch without starting a review, enter:

/observe-report

It lists signal counts, loaded skills, remedy outcomes of applied fixes, stale amendments, the next /observe pass's estimated size, and the Later count. It calls no model and writes nothing.

Optional remote skills

  • retro, from Matt Pocock's skills, adds retrospective guidance for project-environment findings.
  • correct, from Lauren Tan's pstack, can be offered as /skill:correct after at least two evidenced occurrences of a mistake class. Only you may invoke it; observer approval neither launches its repairs nor authorizes commits.

Install these skills separately. Ordinary observation works without them. See optional references for their behavior and setup for source links.

Defaults and limits

  • Interactive review: /observe requires a main session with UI. Amendment delivery also runs in subagents and headless sessions, but requires the extension to stay loaded.
  • Different targets: upstream and plugin skill files stay untouched. Approved local or managed skill changes can edit those skills directly, and approved project fixes use normal editing tools.
  • Instruction delivery, not guaranteed compliance: amendments add guidance to the model's context. Pattern rules are checked against their supplied samples, not proven correct for every possible output.
  • Existing state: loading can migrate legacy observer-owned files. New proposals follow the review workflow; this is not a filesystem sandbox. See storage and migration.

Documentation

  • Setup covers first use, persistent loading, optional skills, and state locations.
  • Observation workflow explains signals, proposals, project scope, and the session digest.
  • Amendments and replay covers storage, delivery, upstream updates, reference files, and pattern checks.
  • Development and tool contracts covers proposal approval, the ledger, module responsibilities, and development checks.
  • Changelog records package changes.

License

The TypeScript implementation, tests, and project documentation are licensed under MIT. The adapted observer.md instructions remain under CC BY 4.0. See third-party notices for attribution and licence scope.

Origins and inspiration

Adapted from task-observer / One Skill to Rule Them All by Eoghan Henn and contributors, licensed under CC BY 4.0. Its approach of finding skill improvements in real work and user corrections is the starting point. This adaptation uses OMP's session history, an explicit /observe command, and separate skill amendments.

ASSERT, from Microsoft Responsible AI, inspired the feedback checks: proposals cite session evidence, pattern rules are replayed against both mistake and legitimate examples, and observation passes check whether earlier fixes failed again.

About

OMP extension to improve workflow and adjust external skills

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages