ClawSeatbelt

Imported from cschanhniem/ClawSeatbelt (AGENTS.md). Install upstream with npx skills add cschanhniem/ClawSeatbelt. Copyright stays with the author.

AGENTS.md

Mission

Build ClawSeatbelt into the local-first trust layer that serious OpenClaw operators install before they add anything else. The product should feel like seatbelts: quiet, decisive, intelligible, and hard to outgrow.

This repository exists to produce three things in lockstep:

1. A world-class OpenClaw plugin.
2. A rigorous operating corpus: plans, task logs, diagrams, threat framing, and launch intelligence.
3. A handoff surface so any new agent can enter cold and still move with precision.

First Read

Start here, in order:

1. clawshield-research.md
2. competitive.md
3. plan.md
4. docs/product/positioning.md
5. docs/architecture/system-overview.md
6. docs/architecture/trust-loop.md
7. docs/architecture/activation-brief.md
8. docs/architecture/share-export-system.md
9. docs/architecture/proof-pack-system.md
10. docs/architecture/compounding-moat.md
11. docs/architecture/default-answer-engine.md
12. docs/architecture/trust-challenge.md
13. docs/architecture/benchmark-harness.md
14. docs/architecture/openclaw-lab-verifier.md
15. docs/architecture/competitor-lab.md
16. docs/product/quickstart.md
17. docs/benchmarks/competitor-artifact-benchmark.md
18. Latest task log in tasks
19. docs/architecture/local-deploy.md
20. marketing/no1-choice-plan.md

Non-Negotiables

• Keep the core local-first. Network use must be optional, explicit, and off the hot path.
• Do not claim prompt injection is solved. Frame the product as risk reduction, visibility, and safer defaults.
• Prefer composition with OpenClaw primitives over reinvention.
• Optimize for operator trust: minimal dependencies, transparent behavior, deterministic modes, and explainable output.
• Growth must come from useful, share-safe artifacts. Never use dark patterns, nagware, or hidden distribution mechanics.
• Write docs and code together. If architecture changes, update diagrams and operating docs in the same pass.
• Break large files before they become a liability.

Repo Shape

• src/core: orchestration, policy evaluation, plugin-facing services.
• src/rules: risk heuristics and redaction rules.
• src/scanner: skill and bundle inspection.
• src/reporting: posture summaries and operator-facing output.
• src/types: shared contracts.
• src/openclaw.ts: OpenClaw-facing plugin entry that matches the official api.on(...) and registerCommand(...) surface.
• docs/architecture: state machines, sequence diagrams, data flow, and system decisions.
• docs/product: positioning, differentiation, and launch logic.
• docs/benchmarks: competitor artifact and integration benchmarks.
• marketing: message architecture, launch sequencing, and default-choice plan.
• tasks: dated work logs.
• test: focused fixtures and regression coverage.

Working Rules

• Before substantial edits, read the latest task log and plan.md.
• Add a new task log for each meaningful work session using tasks/YYYY-MM-DD-short-topic.md.
• Update plan.md when status, sequencing, or scope changes.
• Update SKILL.md when workflow, guardrails, or repo conventions evolve.
• If a change affects adoption, recommendation, or operator sharing, update docs/architecture/trust-loop.md.
• If a change affects install silence, first-reply activation, or the operator's first proof path, update docs/architecture/activation-brief.md.
• If a change affects export rendering, copy structure, or share safety, update docs/architecture/share-export-system.md.
• If a change affects multi-artifact exports, support-thread flows, or recommendation packets, update docs/architecture/proof-pack-system.md.
• If a change affects corpus growth, public proof loops, or trust-safe measurement, update docs/architecture/compounding-moat.md.
• If a change affects support-thread answers, recommendation copy, or maintainer handoff flows, update docs/architecture/default-answer-engine.md.
• If a change affects first-proof installs or challenge flows, update docs/architecture/trust-challenge.md.
• If a change affects corpus benchmarking or package-comparison evidence, update docs/architecture/benchmark-harness.md.
• If a change affects tarball trust, disposable install verification, or allowlist pinning flows, update docs/architecture/openclaw-lab-verifier.md.
• If a change affects live competitor comparisons or category-proof evidence, update docs/architecture/competitor-lab.md.
• If a change affects local OpenClaw install loops, repo linking, or tarball rehearsal flows, update docs/architecture/local-deploy.md.
• If a change affects install, upgrade, or release-verification guidance, update README.md, docs/product/quickstart.md, and docs/release/publish-playbook.md.
• When you add a subsystem, add or update:
  • a state machine diagram
  • a sequence diagram if requests cross component boundaries
  • a data-flow view if data movement or trust boundaries matter

Product North Star

ClawSeatbelt wins if a cautious OpenClaw operator can install it in minutes and immediately gain:

• a shareable posture report
• message risk scoring before the agent sees risky content
• transcript hygiene for persisted tool output
• local skill inspection with concrete remediation
• a coherent explanation of how built-in OpenClaw controls should be configured
• a share-safe artifact worth forwarding to the next operator

Current Strategy

Differentiate on:

• local-first operation
• no-account baseline
• strong operator UX
• skill supply-chain inspection
• unified posture reporting that composes with OpenClaw security audit, pairing, exec approvals, and tool policy
• benchmark proof before marketing claims
• packaging and provenance that make the trust story visible
• an ethical built-in growth loop through trust receipts, skill memos, and benchmark artifacts
• optional interop with heavier policy systems without pulling them into the baseline

Win the choice architecture, not raw feature count. Do not drift into a vague "AI firewall." Stay specific, composable, and useful.