Harness

Skill atlax-tech/harness-armor/skills/harness

Self-improving, tool-first Harness Engineering Agent Skills for Claude Code, Codex, Cursor and TRAE. Build, audit, update and verify AGENTS.md-based repositories.

Install

npx -y skills add atlax-tech/harness-armor --skill harness

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

3 things to look at

26 days oldThe repository was created 26 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Classify the current repository as empty, docs-only, legacy code, managed Harness, custom Harness, or mixed/conflicted; show evidence and route to the correct Harness Armor specialist. Use when the user asks where to start, invokes harness, or wants repository-state triage. Do not use for implementing a known init/build/promotion/update/check/prompt workflow or for ordinary test harnesses.

The file declares its own license as CC-BY-NC-4.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

4.1 KB, as published. Nobody here has run it

Harness router

Act as the read-only entry point. Classify first, explain the evidence, and handoff once. Do not implement a specialist workflow here.

Inputs

Repository root; default to the host's current repository.
Optional stated intent such as audit, sync, initialize, promote, or generate prompts.
No configuration is required.

Load the contract

Treat the directory containing this file as SKILL_ROOT.
Read references/routing.md.
Resolve the shared root in this order:
- SKILL_ROOT/../../shared for source or plugin layout;
- SKILL_ROOT/../.harness-armor for a direct client installation.
Read spec/harness-engineering-v1.md from the resolved shared root. If neither layout exists, stop and report an incomplete installation; recommend npx harness-armor doctor.

Workflow

Read applicable AGENTS.md files without modifying them.
Run python scripts/detect_repository_state.py <root> when Python is available. Otherwise perform the same bounded, read-only inventory with host file/search tools and say the helper was unavailable.
Inspect the reported evidence. The script is a candidate classifier, not a semantic authority. Read the minimum files needed to confirm:
- whether substantive product documents exist;
- whether business code exists;
- whether .harness/manifest.json is valid;
- whether custom agent guidance is coherent;
- whether documents contain unresolved conflicts.
Select exactly one state. Keep MIXED_OR_CONFLICTED when evidence is ambiguous or contradictory; do not resolve conflicts for routing convenience.
If the user already named a valid specialist intent, hand off directly after confirming applicability. Do not ask them to repeat the request.

Otherwise route with this table:

State	Specialist
`EMPTY`	`harness-init`
`DOCS_ONLY`	`harness-build`
`LEGACY_CODE`	`harness-promotion`
`MANAGED_HARNESS`	`harness-update` for explicit sync intent; otherwise `harness-check`
`CUSTOM_HARNESS`	`harness-check`
`MIXED_OR_CONFLICTED`	`harness-check` in read-only diagnostic mode

Use a host-native Skill handoff when supported. Otherwise print the exact supported invocation and stop. Do not copy the specialist instructions into this run.

Authorization and file boundary

This skill is always read-only. It may run bundled read-only scripts. It may not create, edit, delete, rename, install, or normalize repository files. A request to "fix" the result belongs to the routed specialist and its authorization gate.

Forbidden behavior

Do not infer repository state from a directory name or README alone.
Do not call unknown repository scripts.
Do not treat a file count as architecture evidence.
Do not implement initialization, promotion, updates, checks, or prompts.
Do not promise that a specialist ran when only a route was recommended.

Validation and failures

Confirm every state claim has at least one file-level reason.
Report detector truncation, unreadable paths, missing Python, invalid state files, and semantic uncertainty.
If state remains ambiguous, return MIXED_OR_CONFLICTED and route to a read-only check.

Result format

Use assets/state-report.md. Include state, confidence, evidence, uncertainties, selected specialist, invocation, and the fact that no files changed. After a specialist later completes, recommend at most one next step based on its actual result.