Skip to content

refactor: simplify harness setup to one instruction-only skill - #2

Merged
pftg merged 1 commit into
masterfrom
feat/harness-setup-instructions-only
Sep 9, 2026
Merged

refactor: simplify harness setup to one instruction-only skill#2
pftg merged 1 commit into
masterfrom
feat/harness-setup-instructions-only

Conversation

@pftg

@pftg pftg commented Sep 9, 2026

Copy link
Copy Markdown
Member

Summary

Follow-up to #1, implementing the request for an AI-instructions-only plugin without Python
or redundant authoring helpers.

  • Bump harness-setup to 0.2.0.
  • Keep one explicit entry point: /harness-setup:setup.
  • Remove the Python scanner/updater, all Python tests, example plan and Python-specific CI.
  • Remove all four custom agents, including the duplicate skill-creator, and the separate verify skill.
  • Reuse an available skill-creator by its discovered name for instruction-only drafting/review.
    If unavailable, return a creation brief and todo instead of installing or building a duplicate.
  • Fold discovery, clarification, existing-tool selection, review, verification and prioritized
    todos into one workflow.
  • Keep optional approved native edits in one short natural-language reference.
  • Preserve other plugins and unrelated marketplace metadata.

Behavior and boundaries

The package contains six files: one skill, one optional reference, manifest, README, changelog
and license. There are no bundled scripts, agents, hooks, MCP servers or replacement validators.

Assessment is read-only by default. Eligible project-owned edits require review of the exact
candidate, explicit approval and a fresh baseline check. Preserve unrelated settings and every
existing permission protection; global/shared configuration, model/provider routes, hooks and MCP
definitions remain owner tasks.

These are AI instructions, not an enforcement layer. No automatic backups, transactional apply,
authenticated approval mechanism or guaranteed rollback is claimed. Reused skills must not run
scripts, installers or code-based evaluations as part of this workflow.

Validation

  • Claude Code 2.1.220 strict plugin manifest validation passes.
  • Native discovery reports one skill and zero agents, hooks, MCP or LSP servers.
  • Independent AI review: GO for a draft PR, no blockers. The suggested extra manual acceptance
    cases were added; this is not a GitHub maintainer approval.
  • Relative documentation links and git diff --check pass.
  • Marketplace validation has exactly the same two pre-existing warnings as the base; unrelated
    metadata is unchanged.
  • The portable Agent Skills check passes on a temporary projection excluding the two Claude-only
    frontmatter fields. The unmodified skill is not claimed to pass that narrower portable schema.
  • No authenticated Claude workflow acceptance is claimed.
  • The old Python CI is intentionally removed, not replaced with another scripting layer.
    Historical 0.1.0 test/CI results do not validate this revision.

Migration and merge gate

This is a breaking simplification: the former updater commands and /harness-setup:verify are
removed. Old bundles and consuming-project settings are not automatically changed. Use a clean
updated plugin copy rather than retaining old scripts from a manual installation.

  • Maintainer reviews the simplified workflow and removal of mechanical controls.
  • Run the README's manual acceptance cases in a disposable authenticated Claude session.

Keep this PR in draft until those gates are satisfied. No merge, plugin installation or live
harness configuration change is requested.

@pftg
pftg marked this pull request as ready for review September 9, 2026 18:53
@pftg
pftg merged commit 00bf504 into master Sep 9, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant