harness-orchestrate

Harness Orchestration Skill

You are executing a quality-gated SOP workflow using the harness MCP server.

Core Loop

For each step in the workflow:

1. Read the step instruction and acceptance criteria from the harness response
2. Plan your approach - identify what artifacts you need to produce
3. Execute the work using your available tools (Read, Write, Edit, Bash, Grep, etc.)
4. Self-assess honestly against each acceptance criterion before submitting
5. Submit via harness_submit_step with:
   • artifacts: list of what you produced (files written, code blocks, analysis text)
   • self_assessment: honest evaluation of how well you met each criterion
6. Evaluate - after submitting, the harness returns an evaluation prompt:
   • Dispatch the harness-reviewer agent with the evaluation prompt from the response
   • The reviewer agent evaluates your work independently (fresh context, no access to your reasoning)
   • Call harness_report_evaluation with the reviewer's JSON response
7. React to verdict:
   • PASS: proceed to the next step
   • FAIL: read the top_3_fixes carefully, address each one specifically, then resubmit

Evaluation Flow (Subagent Mode)

You do work
  -> harness_submit_step (validates, returns evaluation prompt)
  -> You dispatch harness-reviewer agent with the prompt
  -> Reviewer returns structured JSON scores
  -> harness_report_evaluation (processes scores, returns PASS/FAIL)
  -> If PASS: next step. If FAIL: iterate.

The reviewer agent is a SEPARATE entity with a fresh context. It cannot see your chain of thought. This is by design -- self-evaluation is unreliable.

Dispatching the Reviewer

When harness_submit_step returns stage: "awaiting_evaluation", extract the evaluation prompt from data.evaluation_prompt and dispatch it:

Use the Agent tool to launch the harness-reviewer agent with:
  - The system_prompt from data.evaluation_prompt.system_prompt
  - The user_prompt from data.evaluation_prompt.user_prompt
  - Ask it to return ONLY the JSON evaluation

Then take the reviewer's JSON response and call harness_report_evaluation with it.

Quality Gate Dimensions

The reviewer scores on:

• Completeness (30%): Did you address every acceptance criterion?
• Specificity (25%): Does your output reference THIS project, not generic boilerplate?
• Correctness (20%): Are claims accurate? Does code compile? Does logic hold?
• Actionability (15%): Can the next step act on your output without guessing?
• Format Compliance (10%): Does output match the required structure?

Anti-Patterns to Avoid

• Submitting generic/boilerplate content (the reviewer flags "AI slop")
• Restating the instruction as a finding
• Filler phrases like "This ensures a seamless user experience"
• Submitting without checking every acceptance criterion individually