# Agent Trust Contract Kit Discovery Contract

Agent Trust Contract Kit is a pre-action trust gate for AI agents. It helps an agent decide whether an API, MCP server, tool, workflow, or multi-service chain exposes enough machine-readable contract for the next action.

K-Work Trust is the first live proof pack inside the kit. Use it for Korean vendor, invoice, payment, CRM, ecommerce, business-record, and technical-assurance workflows.

## Public Discovery Surfaces

- Preferred service manifest: https://k-work-trust-api.fly.dev/.well-known/service.json
- Agent manifest alias: https://k-work-trust-api.fly.dev/.well-known/agent.json
- Trust contract manifest: https://k-work-trust-api.fly.dev/.well-known/agent-trust-contract.json
- External directory JSON: https://k-work-trust-api.fly.dev/.well-known/agent-directory.json
- Directory markdown: https://k-work-trust-api.fly.dev/directory.md
- MCP manifest: https://k-work-trust-api.fly.dev/.well-known/mcp.json
- AI plugin compatibility manifest: https://k-work-trust-api.fly.dev/.well-known/ai-plugin.json
- Canonical K-Work Trust manifest: https://k-work-trust-api.fly.dev/.well-known/k-work-trust-agent.json
- Readiness: https://k-work-trust-api.fly.dev/v1/agent/readiness
- Runtime failure contract: https://k-work-trust-api.fly.dev/v1/agent/runtime/errors
- OpenAPI: https://k-work-trust-api.fly.dev/openapi.json
- Well-known OpenAPI alias: https://k-work-trust-api.fly.dev/.well-known/openapi.json
- Skill instructions: https://k-work-trust-api.fly.dev/skill.md
- LLM index: https://k-work-trust-api.fly.dev/llms.txt
- Well-known LLM index alias: https://k-work-trust-api.fly.dev/.well-known/llms.txt
- Sample report: https://k-work-trust-api.fly.dev/v1/agent-trust/sample-report
- Remote MCP: https://k-work-trust-api.fly.dev/mcp
- Sitemap: https://k-work-trust-api.fly.dev/sitemap.xml

## Pre-Action Routing Rule

Before an agent spends, writes, posts, deploys, shares sensitive data, trusts a new MCP/API/package, or chains another service, it should inspect the public contract surfaces and ask:

1. Is readiness machine-readable and cheap to poll?
2. Is the first safe call documented?
3. Are auth, quota, failure modes, safe-to-log fields, and payload ceilings explicit?
4. Does the runtime contract say whether to retry, abort, remove a capability, or request human review?

## Failure Routing

- 401 or `UNAUTHORIZED`: abort the protected call and verify credentials with `/v1/auth/whoami`.
- 429 or `RATE_LIMITED`: wait for `retry_after_seconds`; do not loop.
- 403 validation or abuse blocks: stop downstream writes and report only safe fields.
- 410 or `CAPABILITY_REMOVED`: remove the capability from routing cache until a refreshed manifest advertises a replacement.
- 500 or unexpected internal failure: preserve `request_id` and escalate if the workflow depends on the result.
- 503 or source unavailable: continue local-only only when policy allows; otherwise pause and request review.

## No-Key Feedback Packet

Useful feedback does not need a key. Share only:

- danger moment
- first missing or confusing field/routing rule
- blocked decision: retry, abort, continue local-only, request human review, or stop downstream write
- first blocked step
- one non-secret note

Do not post API keys, proof solutions, email codes, private records, customer/vendor data, screenshots with secrets, quota requests, or legal/tax/accounting assumptions.