harness-ops

Analyzes tasks to propose and execute optimal orchestration patterns — orchestrates sequential/parallel/team/ralph-loop agents, then runs a post-execution verification gate for spec-originated work

Harness maturity assessment — evaluates Claude Code harness using a 6-axis 24-item checklist and 2x3 analysis matrix

Audits all context documents Claude loads (CLAUDE.md, MEMORY.md, skills, agents, plugins) for outdated claims, contradictions, and risky wording

Design and build agent team architecture (delegates to harness-factory). Run when the user asks to 'generate team', 'build agent team', 'build agent pod', 'set up agent architecture', or 'create agent team'. If harness-factory is installed, reads its SKILL.md and executes it; otherwise shows an install guide.

Run a task as a supervised verification loop — 3 gates (Pass/Fail, Quantitative, Qualitative), iterate until they pass, escalate at autonomy boundaries

Reads CLAUDE.md and .claude/rules/* for a project and evaluates quality via LLM-judgment: length, internal contradictions, ambiguities, placeholder content, progressive disclosure, sensitive file protection. Returns CONTEXT_REPORT JSON. Read-only.

Audits a project's automation and verification posture: test infrastructure, formatter/linter PostToolUse hooks, PreToolUse dangerous-action blocks, verifier-agent separation, and whether registered project skills/hooks are actually invoked in recent sessions. Returns AUTOMATION_REPORT JSON.

Analyzes Claude Code session JSONL files to extract execution patterns: plan-ratio, delegation, parallel usage, handoff, repeated n-grams, tool frequency. Uses jq for efficient extraction of tool_use metadata only — never reads prompt text. Scales via split+parallel when many sessions exist. Returns SESSION_REPORT JSON.

Analyzes installed skills/plugins/MCP servers against ~/.claude.json to detect dead, ghost, duplicate, and collision-prone entries AND report the current accessible state (enabled plugins, runtime skills, connected MCP servers). Returns a structured JSON report. Read-only. Use from /check-harness.

Analyze the user's task and propose the optimal orchestration pattern, then execute it. 4 patterns: Sequential Pipeline, Parallel Subagent, Team Mode, Ralph Loop. Situation-aware pattern selection with user confirmation before execution. Use when: "/agent-orchestrate", "agent-orchestrate", "orchestration", "which pattern", "run in parallel", "run sequentially", "team mode", "agent pattern", "suggest execution pattern", "how should we run this", "pick a pattern". Also trigger when the user describes a complex multi-step task that would clearly benefit from agent coordination — e.g., "analyze companies A, B, C", "design, implement, and review", "do these in order", "run 3 at once", or any task with 3+ subtasks where choosing the right execution pattern matters for efficiency.

Harness maturity diagnosis — evaluates the harness cycle (Scaffolding → Context → Planning → Execution → Verification → Compounding) using a **6-axis 24-item checklist** and a **2×3 analysis matrix** (Static/Behavioral/Growth × User/Project). All judgments stem from the gap between "what is set up (Static) ↔ what is actually done (Behavioral)" or "whether the harness is growing (Growth)". Runs 4 subagents in parallel (skill-portfolio-analyzer, session-pattern-analyzer, context-quality-reviewer, project-automation-auditor). session-pattern-analyzer is run twice — once for User global scope and once for the current project — to separate User/Project scopes. Use whenever the user asks to audit their Claude Code harness, review skill portfolio health, evaluate execution patterns across sessions, check project context/rules quality, or wants to know what's missing in their AI setup — even if they don't say "check-harness" explicitly. Trigger: "/check-harness", "check harness", "harness check", "harness audit", "settings check", "what's missing", "harness diagnosis", "maturity check", "my claude setup", "skill cleanup".

Use this skill when the user wants to audit the memory and documents Claude Code loads into context — CLAUDE.md (user global + project + nested), MEMORY.md, @imports, .claude/skills, .claude/agents, .claude/commands, installed plugins — and detect three kinds of issues: outdated claims, mutually contradictory statements, and risky-or-ambiguous wording. Produces a prioritized improvement list at `.drift-reports/`. Zero config. Trigger phrases: "doc drift", "memory drift", "memory audit", "context drift", "docs audit", "document review", "document audit", "memory check", "outdated docs", "document conflict".

Run a task as a supervised verification loop instead of a one-shot prompt. Establishes a loop contract (3 gates: Pass/Fail, Quantitative, Qualitative), then iterates Work → Verify → Fix until every gate passes — emitting an objective evidence report before declaring done. Stops and escalates to a human when an autonomy boundary is crossed (schema change, data-loss migration, auth/payment/security, or a change that conflicts with the spec). Implements the "Ralph loop" technique — the iterative-refinement pattern that agent-orchestrate selects as its Loop pattern. Use when: "/loop", "loop", "run until it passes", "iterate until tests pass", "supervise this until done", "loop.md", "verification loop", "ralph loop", "don't stop until the gates pass", "keep going until criteria met".

Systematically QA test any application — web apps, native macOS apps, Electron apps, CLI tools, interactive REPLs, or anything on screen. Three modes: browser (agent-browser/Playwright, fast, DOM-level), computer (MCP computer-use, screenshot + pixel clicks, any app), and cli (tmux, send-keys + capture-pane for interactive terminals). Auto-selects mode or accepts --browser / --computer / --cli override. Use when asked to "qa", "QA", "test this site", "test this app", "find bugs", "test and fix", "fix what's broken", "dogfood", "exploratory test", "bug hunt", "QA this app", "site test", "app test", "browser QA", "screen test", "native app test". Three tiers: Quick (critical/high only), Standard (+ medium), Exhaustive (+ cosmetic). Produces before/after health scores, fix evidence, and a ship-readiness summary.