ai-harness
AGENTS.md2 starsChanged 30 days agoMITView on GitHub
Version source: HARNESS_VERSION; require agreement with harness/manifest.yaml and POLICY_INDEX.yaml. Read CORE.md, ROUTER.md and POLICY_INDEX.yaml first. Preserve Layer A (source), Layer B (PC shared runtime), Layer C (project bindings). Select only current Task/Phase/Risk/Boundary policies; never preload P01–P26 or all Skills. Jev remains disabled by default, advisory only, with existing-router fallback. Harness upgrade invariant: install/update/upgrade/version/latest-Harness requests route to harness/runtime/harness-update-router.md and harness/update/UPDATE_PROTOCOL.md. Detect current/target version, project mode, development state, custom rules/Skills/policies, checkpoint, compatibility/gap, transition manifest, migration, harness validation, regression, critical backfill, transition audit,…
- Reads credentials
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository5 other files it gives its agents.
- dashboard-builderSkill
- office-docxSkill
- office-pdfSkill
- office-pptxSkill
- office-xlsxSkill
The file
We split this AGENTS.md into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 4
Read CORE.md, ROUTER.md and POLICY_INDEX.yaml first.
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AI Harness runtime entry
Version source: HARNESS_VERSION; require agreement with harness/manifest.yaml and POLICY_INDEX.yaml.
Read CORE.md, ROUTER.md and POLICY_INDEX.yaml first. Preserve Layer A (source), Layer B (PC shared runtime), Layer C (project bindings). Select only current Task/Phase/Risk/Boundary policies; never preload P01–P26 or all Skills. Jev remains disabled by default, advisory only, with existing-router fallback.
Harness upgrade invariant: install/update/upgrade/version/latest-Harness requests route to harness/runtime/harness-update-router.md and harness/update/UPDATE_PROTOCOL.md. Detect current/target version, project mode, development state, custom rules/Skills/policies, checkpoint, compatibility/gap, transition manifest, migration, harness validation, regression, critical backfill, transition audit, resume and benchmark. Never directly replace rules inside an active implementation phase for a major update. Preserve existing behavior and unmanaged project instructions; no retroactive refactor without an explicit migration decision.
For product work route to harness/runtime/orchestrator.md; set acceptance before implementation, retain phase gates, decision gates and beginner briefing, and require execution evidence before PASS. Resolve actual runtime model choices by economy/balanced/frontier tier; logically separate Builder and Reviewer. Do not fabricate commands/results, ignore failed gates or critical findings, weaken verification for token savings, silently break contracts, or load every document.
Managed Codex CLI executions may use scripts/harness-guard-v6-2.mjs for deterministic preflight, token telemetry and final verification. Plain interactive codex sessions are not controlled by the guard; never report guaranteed enforcement without a verified launcher or trusted hook. Only run the guard from a verified pinned source and avoid treating a proposed model tier or JIT file list as observed model selection or actual file reads.
Jev integration (v6.2 candidate): For Jev onboarding/status use scripts/jev-runtime-v6-2.mjs. API keys live only in AI_GATEWAY_API_KEY environment, never in repository or project state. A live synthetic canary and explicit per-project activate --live true enable Shadow; the managed Codex guard may then use Jev on low-risk non-sensitive synthetic metadata only. Jev never overrides the deterministic route, security, deployment, budgeting, review or permissions. Direct interactive Codex is not intercepted. Report disabled/missing key/fallback accurately.
If asked how to access this AI Harness GitHub repository from CLI, how to checkout/review the v6.2 branch, or how to verify Draft PR #3 before merge, JIT-read docs/v6_2/CLI_ACCESS_AND_OPERATIONS_GUIDE.md. Do NOT preload this long operations guide for ordinary coding. It is a runbook, not permission to merge, tag, delete branches, call paid APIs, or install an unreleased source.
Roblox development (v7.0 — JIT ONLY, OPT-IN)
When and only when the task concerns Roblox Studio, game Luau or Studio MCP, read docs/v7_0/ROBLOX_MCP_SETUP.md, harness/capabilities/roblox/capability.json and harness/capabilities/roblox/tool-policy.json as needed. Use the Roblox-published official Studio Agent Plugin and discover the current Studio skill tool; do not falsely label original Harness instructions as Roblox official. Select the exact studio_id before actions; distinguish client registration from real connection and human-submitted evidence from independently observed behavior. NEVER call mutating Studio tools without scoped user authorization, a backed-up test place, and source reconciliation. Keep the published 6.2.0 release and all 26 policies unchanged; do not activate Roblox for non-Roblox tasks. Project Eternal-specific design rules live in docs/v7_0/PROJECT_ETERNAL_REQUIREMENTS.md and must NOT leak into other Roblox projects.
For requests to install/use Roblox Studio MCP from a CLI or prepare the v7.0 release, JIT-read docs/v7_0/INSTALLATION_AND_OPERATIONS.md and docs/v7_0/VERIFICATION_MATRIX.md. Do not call offline CI success a completed Studio Playtest, and do not merge v7.0 or publish v7.0.0 while real user-local Studio E2E evidence remains unavailable. Only the user can authorize changes to their actual Place. Never upload private Studio session IDs or project secrets to PR artifacts.
Roblox v7.0 release decisions must check docs/v7_0/FINAL_RELEASE_GO_NO_GO.md in addition to offline CI; real Studio E2E and original user-workstation evidence are mandatory before 7.0 main merge/tag. Keep this constraint active even if all GitHub Actions checks are green.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

