generated: '2026-09-17'
method: searched
source: >-
  https://docs.aws.amazon.com/lexv2/latest/dg/test-workbench.html,
  https://docs.aws.amazon.com/lexv2/latest/dg/quotas.html,
  https://docs.aws.amazon.com/lexv2/latest/dg/versions-aliases.html — shapes cross-checked against
  smithy/amazon-lex-models-v2-2020-08-07.json
provider: Amazon Lex
providerId: amazon-lex
available: true
kind: live-account test alias + replay harness
note: >-
  Amazon Lex has no separate sandbox account, no test-mode credential and no test/live key prefix pair.
  Every call is a real call against a real AWS account and is billed. What it does ship is two genuine
  rehearsal surfaces, both inside the production account: a reserved draft alias for interactive
  testing, and a replay harness that runs a recorded test set against a bot and reports discrepancies.
  There are no published magic values, test cards or fixture identifiers to record — and none were
  invented.
test_mode:
  separate_credentials: false
  key_prefixes: null
  billed: true
  note: Test traffic is metered exactly like production traffic; see plans/amazon-lex-plans-pricing.yml.
surfaces:
- name: TestBotAlias (TSTALIASID)
  kind: reserved draft alias
  description: >-
    Every bot is created with a built-in test alias pointing at the Draft version, used by the console
    test window and callable from the Runtime V2 API.
  alias_id: TSTALIASID
  docs: https://docs.aws.amazon.com/lexv2/latest/dg/versions-aliases.html
  constraints: >-
    Hard-capped at 2 concurrent conversations on every dimension — text mode, voice mode and session
    management alike — and not adjustable. This is the single most common false signal in Lex load
    testing: an agent benchmarking against TSTALIASID throttles at 2 and reads a 50/125/400-capacity
    service as a 2-capacity one. See rate-limits/amazon-lex-rate-limits.yml.
- name: Test Workbench
  kind: replay harness
  description: >-
    Build a test set from recorded conversations or a manifest, run it against a bot alias, and read
    per-utterance results plus an intent/slot discrepancy report.
  docs: https://docs.aws.amazon.com/lexv2/latest/dg/test-workbench.html
  operations:
  - CreateTestSet
  - DescribeTestSet
  - UpdateTestSet
  - DeleteTestSet
  - StartTestExecution
  - DescribeTestExecution
  - ListTestExecutionResultItems
  - CreateTestSetDiscrepancyReport
  - DescribeTestSetDiscrepancyReport
  - StartTestSetGeneration
  - DescribeTestSetGeneration
  - ListTestSets
  - ListTestExecutions
  since: '2023-06-06'
  since_source: https://docs.aws.amazon.com/lexv2/latest/dg/doc-history.html
- name: Draft bot version
  kind: mutable working copy
  description: >-
    Every bot carries a mutable DRAFT version used for development; numbered versions are immutable
    snapshots created from it with CreateBotVersion.
  docs: https://docs.aws.amazon.com/lexv2/latest/dg/versions-aliases.html
  agent_note: >-
    DRAFT is the closest thing Lex has to a safe workspace. An agent making changes should work in
    DRAFT and promote with CreateBotVersion — that snapshot is also the only rollback material the
    service offers (see conventions/amazon-lex-conventions.yml, reversibility).
test_values:
  published: false
  note: >-
    No magic numbers, test cards, simulator identifiers or hosted test tokens are published for Amazon
    Lex. Recorded as an honest absence rather than left blank.