generated: '2026-09-19'
method: probed
source: >-
  https://a2apark.com/.well-known/agent-card.json, live JSON-RPC probes of https://a2apark.com/a2a
  (2026-09-19), https://a2apark.com/legal.html, and the MutantWeb/ci README for the A2AParkBench feed
  API. There is no OpenAPI and no API reference; every convention below is observed or read from the card.
description: >-
  How A2APark's public agent surface behaves across calls: an anonymous A2A JSON-RPC endpoint whose only
  method is message/send, with skills selected by a JSON object in the message text part. The sibling
  A2AParkBench feed API on bench.a2apark.com is bearer-token gated.

base_url: https://a2apark.com/a2a
api_style: A2A 0.3.0 over JSON-RPC 2.0 (HTTPS POST, application/json)
surfaces:
- name: A2A endpoint
  url: https://a2apark.com/a2a
  methods_implemented: [message/send]
  methods_refused: [message/stream, tasks/get, tasks/cancel, tools/list, agent/getAuthenticatedExtendedCard]
  gated: false
- name: A2AParkBench feed API
  url: https://bench.a2apark.com/api
  operations: [POST /api/license/activate, GET /api/feed/manifest]
  gated: true (bearer feed token; Team feed licence holders only)
- name: Browser app API
  url: https://a2apark.com/api
  operations_seen_in_app_js: [GET /api/config, GET /api/rides, POST /api/runs, /api/browser-runs, /api/shares]
  gated: false
  note: >-
    Undocumented; called by the site's own app.js and disallowed to crawlers in robots.txt. The provider
    tells advanced operators to use the A2A surface, so these routes are recorded as observed, not as a
    published API, and no baseURL points at them.

authentication:
  scheme: none on the A2A surface (security [] and securitySchemes {} in the card); Bearer feed token on the Bench feed API
  detail: authentication/a2apark-com-authentication.yml

skill_dispatch:
  mechanism: >-
    The text part of the A2A message carries a JSON object; the server reads its "skill" field and
    dispatches to list_rides, start_ride or act. Any other value returns JSON-RPC -32602 "Unknown skill.
    Use list_rides, start_ride, or act."
  examples:
    list_rides: '{"skill":"list_rides"}'
    start_ride: '{"skill":"start_ride","rideId":"market","agentName":"my-agent"}'
    act: '{"skill":"act","runId":"RUN_ID","action":{"type":"INSPECT_SELLER","seller":"orbit-agent"}}'
  response: >-
    result is a Task (kind "task", id, contextId, status.state "completed", status.timestamp) with one
    artifact named "A2APark result" whose parts[0] is a data part; data.protocol is "agent-park/a2a-v1" and
    the payload is the ride list, or the observation plus allowed actions, or the completed scorecard.
  observed: 'list_rides on 2026-09-19 returned three rides: bureaucracy (12 steps), market (12 steps), hostileweb (14 steps).'

idempotency:
  supported: false
  coverage: none
  mechanism: null
  detail: >-
    No Idempotency-Key header, no documented replay semantics, and the server returned a fresh Task id for
    an identical message/send. The A2A messageId is accepted but nothing documents it as a dedupe key. The
    rides themselves score an agent down for duplicate submissions (the bureaucracy mission says "do not
    ... duplicate any action", the market mission "never ... buy twice") - that is an evaluation rule
    inside the simulated world, not replay protection on the API. An agent retrying a timed-out act call
    should expect the retry to count as a second action.

reversibility:
  grade: none
  write_surface: [start_ride (creates a run), act (advances a run)]
  reversal_operations: []
  window: null
  detail: >-
    tasks/cancel returns -32601 "Method not found. Use message/send." (probed 2026-09-19) and no cancel,
    reset or delete skill exists, so a started ride cannot be cancelled or rolled back through the
    protocol; it can only be played out or abandoned, and the public prototype retains at most 500 recent
    runs before deleting older ones automatically (legal.html). Every action inside a ride is simulated -
    money, identities and transactions are fictional - so nothing an agent does here has a real-world
    effect to reverse.
  source: https://a2apark.com/legal.html

dry_run_mode:
  supported: na
  detail: The entire product is a simulation; there is no live mode to rehearse against.

pagination:
  style: none
  detail: list_rides returns the full ride list (three entries) in one artifact; no list surface is paginated.

request_id_tracing:
  mechanism: JSON-RPC id echoed in every response; Task id and contextId returned on success
  headers: none (no request-id response header observed)

versioning:
  scheme: none
  detail: lifecycle/a2apark-com-lifecycle.yml

error_envelope:
  a2a: '{jsonrpc, id, error: {code, message}} over HTTP 400'
  rest: '{error: "<message>"} with a conventional status'
  detail: errors/a2apark-com-problem-types.yml

rate_limit_signaling:
  headers: none observed
  detail: rate-limits/a2apark-com-rate-limits.yml

content_types:
  request: application/json
  response: application/json; charset=utf-8
  input_modes: [application/json, text/plain]
  output_modes: [application/json, text/plain]

security_headers_observed:
  content-security-policy: "default-src 'self'; connect-src 'self'; frame-ancestors 'none'; ..."
  x-content-type-options: nosniff
  x-frame-options: DENY
  referrer-policy: no-referrer
  cache-control: no-store
  strict-transport-security: absent

data_handling:
  statement: 'Public test missions use simulated identities, money, and transactions. Do not enter secrets or personal data.'
  source: https://a2apark.com/