Horizon aios dev consistency check
Skill HorizonBrute/Horizon_AI_OS/horizon_system/skills_sbin/horizon_aios_dev_consistency_check
Run an iterative documentation/implementation consistency validation pass against the AIOS consistency-check standard, fixing unambiguous drift and surfacing decisions until 100% clean or blocked on user input. Use when asked to run a consistency pass/check, validate AIOS standards, or verify docs and implementation are consistent (e.g. after a feature lands).From its SKILL.md
npx -y skills add HorizonBrute/Horizon_AI_OS --skill horizon_aios_dev_consistency_checkAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 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.
- 0 stars0 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.
SKILL.md
4.7 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it
Skill: /horizon_aios_dev_consistency_check
Model preference: #midcost (per horizon_aios_model_prefs.md; overridable by a prompt directive).
Run a consistency validation pass over the Horizon AIOS repo using the standing
standard at $HORIZON_DOCS/development_tools/consistency_checks.md. That file —
not this one — is the single source of truth for the checks (IDs CC-*), the
authority hierarchy, the run protocol, and the report format. This skill is the
invocation entry point that runs it.
When to invoke
- "run a consistency pass/check", "validate AIOS standards", "check docs vs implementation are consistent", "verify the AIOS is internally consistent".
- After a feature or docs change lands, to confirm no drift was introduced.
Run this as a subagent (do this first)
A consistency pass reads broadly across docs and code, so running it inline bloats
the invoking session's context. When invoked, delegate the work to a
general-purpose subagent rather than doing it in the main session (this follows
the agent delegation model in agents.md — the main session orchestrates,
subagents do the file-heavy work). Only run inline if subagents are unavailable.
How to delegate:
- Spawn a general-purpose subagent (the harness's Agent/Task tool,
subagent_type: general-purpose). Subagents start cold — give it enough scope and context to be self-sufficient. - Instruct it: "read
$HORIZON_DOCS/development_tools/consistency_checks.mdand run its protocol over <scope>; fix unambiguous drift, collect judgment calls into 'Needs user input', report each pass in the §4 format; do not commit." - The subagent runs the iterate-until-clean-or-blocked loop and returns its report; the main session relays it and brings any "Needs user input" items to the user.
For a full-repo pass you may fan out subagents by check group (A–I) or by area and have the main session synthesize one report, then re-run until clean or blocked.
Execution
- Read
$HORIZON_DOCS/development_tools/consistency_checks.mdin full. - Determine scope from the user's prompt (default: the whole repo; may be narrowed, e.g. "only this session's changes"). State the scope up front.
- Run the protocol in §1 of that file:
- Execute every in-scope check (
CC-*). Gather positive evidence for each verdict —file:line, a command + its output, or a quoted passage. No false greens: anything unverified isUNKNOWN, neverPASS. - Fix unambiguous drift directly (stale paths, renamed flags, missing index entries, one-sided platform implementations).
- Surface judgment calls into a single "Needs user input" list — design trade-offs, authority conflicts not resolved by §2, behavior changes, or anything destructive/outward-facing. Do not guess.
- Re-run a full pass after any fix (a fix can introduce new drift).
- Terminate at 100% clean (all PASS or justified N/A) or when blocked on user input.
- Execute every in-scope check (
- Report each pass in the §4 format. Keep a running tally of pass count and the terminal state.
Notes for the executing agent
- Never weaken or delete a check to make a pass go green. If a check itself is
wrong/outdated, flag it as a finding and ask before editing
consistency_checks.md(§1 rule). - Apply the authority hierarchy (§2) to conflicts:
philosophy ▶ dev_values ▶ architecture_decisions ▶ other docs ▶ implementation. A lower doc that contradicts a higher authority is the defect. - Cross-platform (Group C) and onboarding/offboarding (CC-I3) require enumerating the platform-specific assets and confirming parity — verify it, do not infer it.
- CC-G4 (doc index) and CC-G5 (token economy of context-loaded files) often pair
with
/horizon_aios_documentation_index_update; run or recommend it when CC-G4 fails. - CC-T2 must be delegated to the
/terseness-checkskill — do not inline- evaluate terseness. Spawn it as a sub-agent, collect its report, and map each FAIL finding toCC-T2 FAILin this pass's report. CC-T1 (index currency) can be verified inline by tracing the @-import chain. - Commit fixes only if the user asks (DCO sign-off per repo rules); otherwise leave changes in the working tree and report.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.