harness-ops

Plugin harness-ops from marketplace agent-harness in Marius-Juston/agent-harness. Commands read from the repo's manifest — 13 days ago.

Skills (3)

• harness

  How to change a project's agent surface safely — CLAUDE.md, .claude/ settings and agents, skills, hooks, the [tool.devkit] config, CI — without drift or bloat, and how to measure whether a change helped.

• lead

  The lead session's routine in a repository run with agent-harness — orient, integrate finished PRs, file and rank issues, dispatch workers with one-line prompts, resume after limits or compaction, keep HANDOFF current.

• research

  Recursive deep research in a repository run with agent-harness — decompose a question, fan out researcher subagents in rounds, chase the gaps each round reveals until saturation, then synthesize one cited house-style note, audit it adversarially, and verify every source.

Slash commands (1)

• /harness-ops:gotcha

  Capture a one-line lesson about a path into the .claude/rules file scoped to it, with (#issue, date) provenance

Subagents (7)

• architect

  Read-only consolidation review of the whole repository, run when `uv run dev orient` says a review is due or the owner asks for one. Builds a feature inventory from `uv run dev consolidation report` and files a leaner structure as an epic of small issues for builders. It never edits, moves or delete

• builder

  Implements one build/infra/bug issue test-first in its own worktree and hands back a gated PR. Use for any code change larger than a few lines, one issue per builder; parallel builders only on issues whose Owns do not overlap.

• research-auditor

  Adversarial, read-only audit of one research note. Re-fetches its sources and checks that each claim is supported, labelled correctly and not overstated. Use before a research PR is integrated, or when a note drives a design decision.

• researcher

  Researches one question or one research issue and writes or extends one cited note in docs/research/ (house style, labelled claims, verified sources). Use for any "type:research" issue, for one sub-question of a research round (the harness-ops:research skill), or before any design or algorithm choic

• reviewer

  Read-only pre-merge review of one PR against its issue's DONE WHEN, the research it cites and the project rules. Use for PRs touching the harness, several Owns, or anything a user sees.

• suite-auditor

  Read-only judgement pass of a size-tiered test-suite audit, run when `uv run dev suite-audit report` lists judgement items. Reads the tests behind them and files at most 5 type:test issues. Never edits, moves or deletes a test.

• tester

  Writes adversarial, property-based and regression tests for one issue or one merged change (inputs nobody planned for, wrong state, crashes midway), fixes small defects it exposes, files the rest. Use after merges that add behavior, or for "type:test" issues.

Hooks (10)

• SessionStart

  "${CLAUDE_PLUGIN_ROOT}"/scripts/session_brief.sh

• PreToolUse

  "${CLAUDE_PLUGIN_ROOT}"/scripts/guard_bash.py

• PreToolUse

  "${CLAUDE_PLUGIN_ROOT}"/scripts/uv_add_guard.py

• PreToolUse

  "${CLAUDE_PLUGIN_ROOT}"/scripts/read_guard.py

• PreToolUse

  "${CLAUDE_PLUGIN_ROOT}"/scripts/edit_guard.py

• PreToolUse

  "${CLAUDE_PLUGIN_ROOT}"/scripts/spawn_guard.sh

• SubagentStart

  "${CLAUDE_PLUGIN_ROOT}"/scripts/spawn_guard.sh

• SubagentStop

  "${CLAUDE_PLUGIN_ROOT}"/scripts/completion_gate.py

• PostToolUse

  "${CLAUDE_PLUGIN_ROOT}"/scripts/lint_after_edit.py

• WorktreeCreate

  "${CLAUDE_PLUGIN_ROOT}"/scripts/worktree_create.sh