Component Map

June 18, 2026 ยท View on GitHub

This map connects harness engineering concepts to files in a target repository.

Harness ConceptTarget Repo ArtifactStarter Template
Harness theorypractical model for repository-level agent supportdocs/theory/harness-engineering.md
Agent instructionsAGENTS.mdtemplates/generic/AGENTS.md
Architecture decisionsdocs/decisions/*.mdtemplates/generic/docs/decisions/000-template.md
Failure memorydocs/failures/*.mdtemplates/generic/docs/failures/000-template.md
Coding conventionsdocs/conventions/*.mdtemplates/generic/docs/conventions/coding.md
Domain knowledgedocs/domain/*.mdtemplates/generic/docs/domain/glossary.md
Document drift checkscripts/check_docs_drift.pytemplates/generic/scripts/check_docs_drift.py
Structure drift checkscripts/check_structure.pytemplates/generic/scripts/check_structure.py
Encoding hygiene checkscripts/check_encoding_hygiene.pytemplates/generic/scripts/check_encoding_hygiene.py
Effectiveness plan checkscripts/check_effectiveness_plan.pytemplates/generic/scripts/check_effectiveness_plan.py
Failure-memory checkscripts/check_failure_memory.pytemplates/generic/scripts/check_failure_memory.py
Decision-memory diff warningscripts/check_decision_memory.py, .harness/decision-memory-rules.jsontemplates/generic/scripts/check_decision_memory.py, templates/generic/.harness/decision-memory-rules.json
Harness readiness diagnosticsix-element /harness doctor report with coupling findingscommands/harness-doctor.md, docs/scoring/harness-score-rubric.md
Harness adoption workflow/harness adopt reportcommands/harness-adopt.md, docs/adoption-workflow.md, docs/templates/adoption-report.md
Harness source tracking.harness/source.jsondocumented in commands/harness-update.md
Harness update workflow/harness update reportcommands/harness-update.md
Harness refresh workflow/harness refresh reportcommands/harness-refresh.md
Harness change-set review/harness review or /harness review sub-agent reportcommands/harness-review.md, docs/templates/harness-review-report.md
Universal Agent Skills packageCodex and Claude Code skill/plugin adapters for /harness ... workflowsagent-skills/, agent-skills/.codex-plugin/plugin.json, agent-skills/.claude-plugin/plugin.json, docs/agent-skills-package.md
Agent Skills package validationportable skill, metadata, manifest, and reference drift checkscripts/check_agent_skills_package.py
Baseline harness score scansix-element repository health and coupling scanscripts/harness_doctor.py
Deterministic benchmark tasksrepo-owned benchmark task definitions and project-specific oracles for harness-agent-benchmark-runnerbenchmarks/README.md, benchmarks/tasks/*.json
Optional scheduled harness check.github/workflows/harness-check.ymltemplates/generic/.github/workflows/harness-check.yml
External API work recipeserver-only API boundary, redaction, live/mock fallback, and smoke checksdocs/checklists/external-api-work.md
Decision and failure memory guidanceexamples for when to record ADRs, failure notes, domain docs, or final-report notesdocs/checklists/decision-failure-memory.md
Verification script patternscustom smoke checks and transparent check:harness compositiondocs/checklists/verification-scripts.md
Dogfood evidence adoptionsource tracking, task outcome, failure memory, gate placement, and claim-boundary reviewdocs/checklists/dogfood-evidence-adoption.md
Stack-specific ruleslint/type/pre-commit/framework snippetstemplates/profiles/*
Stack profile guideavailable profiles and how to treat snippets as reference materialdocs/profiles.md
Profile absorptionchecklist for turning profile snippets into project rulesdocs/checklists/profile-absorption.md
Adoption report examplefinal adoption summarydocs/templates/adoption-report.md, examples/*-adoption-report.md
Effectiveness evaluationagent mistake reduction measurementdocs/evaluation.md, docs/templates/effectiveness-report.md
Task outcome recorddocs/effectiveness/task-outcomes/*.yamldocs/templates/task-outcome.yaml
Validation coveragetests, smoke checks, and E2E coverage notesdocs/validation.md
Lifecycle pilot resultsprompt-first adoption behavior evidencedocs/examples/lifecycle-pilot-results.md

Minimum Useful Adoption

For a very small project, the agent should add or adapt only:

  • AGENTS.md
  • docs/decisions/000-template.md
  • docs/failures/000-template.md
  • scripts/check_docs_drift.py
  • scripts/check_structure.py
  • scripts/check_encoding_hygiene.py
  • scripts/check_effectiveness_plan.py
  • scripts/check_failure_memory.py

Then grow the harness as the project and agent usage mature. Use the optional GitHub Actions workflow skeleton with --with-ci only after confirming the target repository uses GitHub Actions.

Use /harness adopt when the maintainer wants first-time prompt-first harness adoption. The command inspects the target, applies only the smallest useful harness pieces, and finishes with an adoption report and effectiveness measurement plan.

Use /harness doctor when the maintainer wants a diagnostic score before or after adoption. The command reports readiness; it does not install harness files.

Use /harness update after adoption when the maintainer wants to refresh the local kit reference and selectively apply new harness guidance. The command records the current kit source in .harness/source.json and reports applied, skipped, and manual-review items.

Use /harness refresh after adoption when the maintainer wants to clean up the target harness itself. The command reviews stale docs, duplicated guidance, obsolete records, and unused checks, then reports keep, update, merge, archive/delete candidate, and manual-review items without deleting files unless the user explicitly approves the specific files.

Use /harness review before finishing a change when the maintainer wants an opposing harness-engineering review of the current diff. The command reports source-of-truth risks, unnecessary automation, missing checks, durable memory gaps, overreach, and follow-up recommendations without modifying files unless the user explicitly asks to apply fixes afterward.

Use /harness review sub-agent when the maintainer explicitly wants the read-only reviewer subagent path. It uses the same report, still modifies no files, and records a fallback reason if the active runtime cannot call a subagent.

Use the Universal Agent Skills package when maintainers want the prompt-first workflows exposed as runtime-native skills. The package keeps this repository's workflow docs as the canonical source while providing bundled fallback references for standalone skill installs.