Skip to content

feat: add agentic-harness plugin (harness-setup + harness-review)#8

Merged
MrBogomips merged 8 commits into
mainfrom
feature/agentic-harness
Jun 1, 2026
Merged

feat: add agentic-harness plugin (harness-setup + harness-review)#8
MrBogomips merged 8 commits into
mainfrom
feature/agentic-harness

Conversation

@MrBogomips

Copy link
Copy Markdown
Owner

Summary

Adds agentic-harness, an MIT plugin for the mrbogomips marketplace that stands up, assesses, and maintains a project-local agentic harness — the agents, skills, and orchestrator under .claude/ that make a repo work well with Claude Code. It is a meta-tool: it builds and maintains the agents/skills that do the domain work, rather than doing that work itself.

Independent, clean-room reimplementation inspired by the open-source harness plugin by revfactory (Apache-2.0); all prose re-authored, anonymized, and in the marketplace house voice. Credit is in the repo README only.

Architecture — two skills on a read/write axis

  • harness-setup (writer): domain analysis → execution-mode + architecture-pattern selection → generate agent definitions → create skills → build orchestrator → register the CLAUDE.md pointer → record change-history → capture feedback. Also extends, applies a review context, and syncs drift.
  • harness-review (read-only): inventories .claude/, reads the CLAUDE.md pointer + change-history, detects drift, and assesses effective usage from a fixed signal set (project auto-memory, the pointer + history, the .claude/ inventory, the tools registry), then emits a prioritized review context that hands off to harness-setup.

The loop is review → setup → review again — one reader, one writer.

Three cross-cutting concept docs live once in agentic-harness/shared/ (harness-model, claude-md-pointer, execution-modes) and are cited via ${CLAUDE_PLUGIN_ROOT}/shared/...; per-skill references live under each skill's references/.

Notable design points

  • Execution modes: agent team is the preferred default when coordination is needed and the experimental team tools are present; a mechanical subagent fallback is documented for when they are not. The team-tools caveat is stated up front, not buried.
  • Optional tool-discovery step (harness-setup Step 1b, on request only): a search-optimized subagent finds candidate MCP servers and plugins fitting the project's domain/stack and checks the local config. Nothing is adopted without explicit per-tool approval. Accepted tools are registered by role in a tools.md under the orchestrator (role → preferred tool → alternative → status); agents and skills reference tools by role and fall back to the alternative when one is unavailable. Periodic tool review: assessed by harness-review, acted on by harness-setup.
  • No commands — deprecated in Claude Code (merged into skills); both skills are slash-invocable as /agentic-harness:harness-setup and /agentic-harness:harness-review.
  • agentic-harness inserted first in marketplace.json (alphabetical).

Test plan

  • bash tests/ci/run-structural-tests.sh → "All test suites passed" (plugin, references, connectors)
  • No ~~ tokens, no CONNECTORS.md required
  • No angle brackets in skill description: lines; both SKILL.md bodies < 500 lines (221 / 134)
  • All per-skill references/ resolve; all shared docs cited via ${CLAUDE_PLUGIN_ROOT}/shared/...
  • Fully anonymized (no real project/person names) and Korean-free
  • Trigger sanity checklist per skill (8 should-trigger + 8 should-NOT each), confirming clean read/write separation:

harness-setup should NOT trigger on assess/audit/usage phrasings (those go to harness-review); harness-review should NOT trigger on build/extend/register/sync/fix phrasings (those go to harness-setup). The one subtle boundary — "is there a better tool?" (assess → review) vs "swap the tool" (act → setup) — is handled by the trigger verb.

On explicit request, harness-setup can dispatch a search-optimized subagent to
find candidate MCP servers and plugins that fit the project, check the local
config for already-available tools, and — only with explicit per-tool approval —
register accepted tools by role in a tools.md registry under the orchestrator
(role, preferred tool, alternative, status). Agents and skills reference tools by
role and fall back to the alternative when one is unavailable. Periodic tool
review is documented in maintenance.md (assess via harness-review, act via setup).
@MrBogomips MrBogomips merged commit aa935c1 into main Jun 1, 2026
1 check passed
@MrBogomips MrBogomips deleted the feature/agentic-harness branch June 1, 2026 13:51
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