generated: '2026-07-21'
method: searched
source: https://docs.lightcone.ai/guides/production + https://docs.lightcone.ai/guides/errors + https://docs.lightcone.ai/guides/authentication + https://docs.lightcone.ai/guides/limits-and-behaviors
authentication:
  style: bearer API key (sk_ prefix), TZAFON_API_KEY env var
  artifact: authentication/tzafon-authentication.yml
idempotency:
  supported: true
  mechanism: idempotency_key body parameter on POST /agent/tasks (task start)
  docs: https://docs.lightcone.ai/guides/production
  behavior: >-
    A retry with the same idempotency_key returns the same task with
    reused: true instead of starting a new one. Derive the key from your own
    job ID so retries at any layer (queue, HTTP client, SDK) deduplicate to
    one task.
pagination:
  documented: false
  notes: List endpoints (GET /computers) return full arrays; no cursor/offset pagination is documented.
request_tracing:
  documented: false
versioning:
  style: model-version pinning + additive optional fields; deprecated fields signaled via Deprecation/Sunset headers
  artifact: lifecycle/tzafon-lifecycle.yml
error_envelope:
  format: RFC 7807 application/problem+json (type, title, status, detail, instance)
  artifact: errors/tzafon-problem-types.yml
rate_limit_signaling:
  status: 429 on plan concurrent-computer limit, with retry guidance
  artifact: rate-limits/tzafon-rate-limits.yml
retries:
  sdk_default: SDKs retry automatically on connection errors and 5xx (default 2 retries)
  guidance: exponential backoff on 429/5xx; branch failed task events on retryable flag
streaming:
  style: Server-Sent Events for task streams; newline-delimited JSON for /exec
  resumption: >-
    Task SSE streams are resumable via the standard Last-Event-ID header or
    ?from=<seq>; a replay_gap event signals evicted events (re-sync from
    retrieve_status). Streams remain available ~10 minutes after a task ends.
coordinates:
  convention: model returns 0-999 normalized coordinates (since April 8, 2026); denormalize to pixels client-side. Direct computer actions take screenshot-pixel coordinates.
timeouts:
  per_action: ~20 seconds server-side default; per-action timeout_ms accepted
  task_ceilings: max_steps and max_duration_seconds should be set on every task
compatibility:
  - OpenAI drop-in compatibility on /chat/completions and /v1/responses (use OpenAI SDK with a Lightcone base_url)
  - Anthropic Messages compatibility (Anthropic-style error bodies)