Skip to content

FAQ.md

Latest commit

 

History

History
55 lines (31 loc) · 3.17 KB

FAQ.md

File metadata and controls

55 lines (31 loc) · 3.17 KB

FAQ

What is this kit?

A copy-paste AI operations starter: genetic tags, navigation map, Cursor rules, optional AgentStack bridge. Version tracks AgentStack platform patch (e.g. 0.4.11).

Windows: PowerShell blocks install.ps1

Use SETUP.cmd or install.cmd — they bypass default ExecutionPolicy. See meta/docs/INSTALL_WINDOWS.md.

Should kit files be in git?

  • Team shares map/genes: commit installed files (--gitignore-kit none, default for --yes CLI).
  • Privacy / local-only AI context: --gitignore-kit full or GITIGNORE=full with install.cmd. Already-tracked files need git rm -r --cached — see docs/ai/OPERATIONS.md in your project.

npx vs cloning the repo

npx @agentstack/genetic-ai-starter init --yes --target ./my-app --profile standard --project-name "My App" --domain app

Requires Node 18+.

Benchmark numbers in README — what do they mean?

Task score (0–10): sum of rubric dimensions per task (correct files, navigation path, scope, outcome, efficiency). Success = score ≥ 6. Median = middle value across 14 tasks — use with success rate and task table T04/T05/T13.

Harness: shop-api, 14 tasks, scorer 1.2.1, synthetic policy transcripts. Run locally:

node benchmarks/scripts/prepare-all-arms.mjs --force
node benchmarks/scripts/run-matrix.mjs

Why agents_md median looks high (8): benchmark arm with optimistic transcripts, not your production agent. Map-first (genetic) is only 7%; arm agents_md_weak (same AGENTS file, pessimistic transcripts) median 2.5. Compare kit + indexes: success 100%, map-first 86%.

What kit improves vs bare: T04 refusal 2→8, T05 maintenance 4→10, T13 release gate, T08 10 with indexes. Primary narrative: AI_RELEASE_AUTONOMY_ru.md. Metrics: METRICS_GLOSSARY.md.

What we do not claim: Cursor/vendor certification — validate on your repo via benchmarks/METHODOLOGY.md § Manual validation.

Can a cheap / weak agent match top models?

On repo-bound engineering tasks — often yes for stability, not for everything. The kit raises the floor: harness arm agents_md_weak scores median 2.5, 0% success; kit + indexes scores median 9, 100% success on the same 14 tasks. That is not “mini = Opus on product design”; it is map + genes + doctor so weak agents stop grep/sed chaos and hit T04/T05/T13 consistently. Full narrative: AGENT_FLOOR.md · RU: AGENT_FLOOR_ru.md. Evidence table: DOC_CLAIMS_AUDIT.md.

Easter eggs?

Some installs include optional community genes and doctor --beacon. See COMMUNITY.md (no full spoilers).

Issues: monorepo or public repo?

Public mirror for community bugs and PRs. Monorepo kit label is for AgentStack maintainers syncing SoT.

Do I need AgentStack hosted?

No. Standard/minimal profiles work offline in Cursor. Full/founder profiles add optional MCP/8DNA documentation overlays.