generated: '2026-09-19'
method: probed
source: >-
  Cross-cutting behaviour observed on live, unauthenticated calls to the three surfaces on
  2026-09-19, plus what the OpenAPI, the A2A card and agents.json state. Where a convention is not
  documented and was not observed, it is recorded as absent rather than assumed.
description: >-
  How RS Performance's REST gateway, A2A endpoint and MCP server behave across operations:
  authentication, idempotency, reversibility, pagination, tracing, versioning, error envelopes and
  rate-limit signalling.
base_urls:
  rest_gateway: https://ai.rsperformance.online
  a2a_jsonrpc: https://rsperformance.online/
  mcp: https://mcp.rs3d.pl/
api_style: >-
  REST over HTTPS with JSON bodies (gateway), JSON-RPC 2.0 over HTTPS plus HTTP+JSON message
  endpoints (A2A), MCP Streamable HTTP with SSE responses (MCP).
authentication:
  scheme: none — every surface is anonymous by declaration and by test
  detail: authentication/rsperformance-online-authentication.yml
idempotency:
  supported: false
  coverage: none
  mechanism: null
  docs: null
  write_surface:
    - A2A message/send and message/stream (create a task; JSON-RPC id is a correlation id, not an idempotency key)
    - MCP write_intake_note (creates a note) and diagnostic_ingest_brand_knowledge (saves artifacts on the VPS)
  note: >-
    No Idempotency-Key header, no client-supplied request id that dedupes, and no documented replay
    behaviour on any write. A retried message/send creates a second task. The REST gateway has no
    writes (semanticSearch is a POST-shaped read), so idempotency is moot there.
reversibility:
  grade: documented
  write_surfaces:
    - surface: A2A task (created by message/send or message/stream)
      reversal_operation: tasks/cancel (JSON-RPC; legacy alias CancelTask)
      reversal_declared_in: a2a/rsperformance-online-agent-card.json supportedInterfaces[3].x_rs_transport_note and the live -32601 method list; the legacy card describes "cancellable short-lived tasks"
      window: not stated — no document says how long after creation a task can be cancelled or what state blocks cancellation
      docs: https://rsperformance.online/.well-known/agent-card.json
    - surface: MCP write_intake_note
      reversal_operation: none — no delete/undo tool in tools/list
      window: n/a
    - surface: MCP diagnostic_ingest_brand_knowledge
      reversal_operation: none — no tool removes a saved knowledge packet
      window: n/a
  note: >-
    One reversal path exists (tasks/cancel) with no stated window, which grades as documented, not
    verified. No refund/void semantics apply — the surface takes no payments; booking is a
    routing/contact skill, not a transaction.
dry_run_mode:
  supported: false
  note: No dry-run, validate-only or sandbox flag on any write.
pagination:
  style: none documented
  rest_gateway: limit parameter only (1-20, default 5) on semanticSearch; no cursor, offset or has_more
  a2a: tasks/list returned 34 tasks in one response with no paging parameters documented
  mcp: limit / top_k / max_sources integer caps per tool; no cursor
field_expansion:
  supported: false
metadata:
  supported: false
request_tracing:
  request_id_header: null
  note: No request-id or trace header observed on any response.
versioning:
  scheme: version numbers inside each document; endpoints unversioned
  detail: lifecycle/rsperformance-online-lifecycle.yml
error_envelope:
  rest_gateway: '{"detail": "..."} (FastAPI), application/json'
  a2a: 'JSON-RPC 2.0 error object {code, message}, HTTP 200'
  mcp: 'JSON-RPC 2.0 error object with id "server-error", real HTTP status (400/406)'
  detail: errors/rsperformance-online-problem-types.yml
rate_limit_signaling:
  headers_observed: [X-RateLimit-Limit, X-RateLimit-Remaining]
  where: 'rsperformance.online only (A2A POST /: 60; GET /api/knowledge/search: 30); none on ai.rsperformance.online or mcp.rs3d.pl'
  exhaustion_status: not observed
  detail: rate-limits/rsperformance-online-rate-limits.yml
caching_guidance:
  agent_card_ttl: '>= 600 s (card x_rs_ai_friendly.rate_limit and security.txt fair-use comment)'
  well_known_ttl: '>= 600 s (security.txt)'
  beacon: 'auto.rs3d.pl/webhook/ai-beacon: Cache-Control public, max-age=300; refreshed every 6 h'
languages:
  priority: [pl-PL, en]
  note: Queries accepted in Polish or English (OpenAPI description, card languages_priority); most content and error prose is Polish.
attribution:
  required: true
  format: 'Source: RS Performance (rsperformance.online)'
  license: content CC-BY-SA-4.0 for AI training per the x-ai-training-policy header and freshness.json; gateway agent.json says attribution is required when citing