ReqManagedAgents. Providers. ClaudeManagedAgents. Federation
(ReqManagedAgents v0.17.0)
Copy Markdown
View Source
Anthropic workload identity federation settings for a Claude Managed Agents
client, and the access-token cache behind them. Build a client with
ReqManagedAgents.Providers.ClaudeManagedAgents.Client.new/1; this module is
its :federation field.
Identity tokens carry a jti and are single use, so every exchange obtains a
new one and only one exchange per credential runs at a time. Only the caller
that performs a refresh waits for it. Every other caller holding a usable cached
token is served it at once, while another process refreshes and during the retry
pause after a failed refresh. A caller with no usable token waits at most
:exchange_timeout. For an access token expiring at e:
- before
e - 120 sthe cached token is served; - from
e - 120 stoe - 30 sone caller exchanges and the others keep the cached token; if the exchange fails the cached token keeps serving while retries pause for 10 seconds; - from
e - 30 sthe cached token is unusable: exchange or fail.
The 120 s and 30 s margins shrink to half and a quarter of the token's lifetime when that is shorter (a 60 s token refreshes from 30 s before expiry).
The identity-token function is given :identity_token_timeout (10 s), and the
exchange request :exchange_connect_timeout (5 s) and :exchange_receive_timeout
(10 s); none uses the client's :receive_timeout. One refresh therefore takes at
most about 25 s, and the :exchange_timeout default (30 s) covers it, so a
waiter outlasts a holder that succeeds. When raising one of the three timeouts,
raise :exchange_timeout to at least their sum.
Tokens are cached per node, keyed by the settings and the identity-token
source (the file path, the function, or :cache_key when given). Each new
key is a separate exchange that spends a single-use identity token. Entries
whose token has expired are dropped whenever a new entry is written.
Summary
Types
What is known of a failed exchange's response: the HTTP status (nil when there was
no response) and Anthropic's request-id header, never the body.
:cache_key is any term; clients sharing it, and the other settings, share one
cached access token even when each passes its own :identity_token closure.
Types
What is known of a failed exchange's response: the HTTP status (nil when there was
no response) and Anthropic's request-id header, never the body.
@type reason() :: :identity_token_unreadable | :identity_token_invalid | :identity_token_unavailable | :identity_token_timeout | :exchange_timeout | {:exchange_rejected | :exchange_failed | :exchange_invalid_response, http_detail()}
@type t() :: %ReqManagedAgents.Providers.ClaudeManagedAgents.Federation{ cache_key: term(), connect_timeout: pos_integer(), exchange_timeout: pos_integer(), identity_token: identity_token() | nil, identity_token_file: String.t() | nil, identity_token_timeout: pos_integer(), organization_id: String.t(), receive_timeout: pos_integer(), rule_id: String.t(), service_account_id: String.t() | nil, workspace_id: String.t() | nil }
:cache_key is any term; clients sharing it, and the other settings, share one
cached access token even when each passes its own :identity_token closure.