Headers
Request headers
Section titled “Request headers”| Header | Effect |
|---|---|
Authorization: Bearer <key> |
Managed address: the issued key. Status and metrics: the read token. Control API: the owner token |
X-Prism-Admin-Token |
Owner token for the control API, status, cost, metrics and traces |
X-Prism-Read-Token |
Read-only token (PRISM_READ_TOKEN) for status, cost and metrics |
X-Prism-Strategy |
Strategy for this request: balanced, quality, cheap, fast, private |
X-Prism-Session |
Names the conversation; same as body session_id |
X-Prism-App |
Names the agent; a mapped name selects a routing profile |
X-Prism-Profile |
Selects a routing profile by name |
X-Prism-Allow-Provider-Failover |
true lets a provider::model request go to another provider |
X-Prism-Tool-Compression |
Tool-output compression for this request |
X-Prism-Outcome: ignore |
Served, billed and recorded as usual, but left out of outcome judgment and of the Experiments page’s policy and model tables. For evaluation runs, benchmarks and judges that send one task to several models |
X-Request-Id |
Managed address: your request id (one value) |
See Choosing a model for how the routing headers interact.
Response headers
Section titled “Response headers”Set on answers to Chat Completions and Responses requests, and on the translated dialects where they apply.
| Header | When | Meaning |
|---|---|---|
X-Request-Id |
Every answer | Prism’s id for this request |
X-Prism-Provider |
Routed answer | Provider that served it, after any failover |
X-Prism-Model |
Routed answer | Model id sent to that provider |
X-Prism-Selector, X-Prism-Reason, X-Prism-Quality, X-Prism-Candidates |
Prism chose the model | Selector, why this model, its quality score, number of candidates |
X-Prism-Role-Alias, X-Prism-Resolved-Model |
A role alias was used | Alias name and the model it resolved to |
X-Prism-Attempts, X-Prism-Attempt-Providers, X-Prism-Attempt-Reasons |
Routed answer | Number of attempts, providers tried, result of each |
X-Prism-Cost-USD, X-Prism-Cost-Status, X-Prism-Pricing-Revision |
Routed answer | Cost in USD; known, partial or unknown; price table revision |
X-Prism-Trace-Id, X-Prism-Orchestration, X-Prism-Strategy-Effective, X-Prism-Selection-Reason |
Default configuration ([v2] on) |
Trace id; calls, escalation and cost; strategy used; why the model was chosen (for example tier_floor_order) |
X-Prism-Session-Migrated |
This turn was served by another account or model than the previous turn | Present only then |
X-Prism-Served-By |
[harness] announce_served_model = true |
provider/model |
X-Prism-Adjusted-Params |
Prism changed request parameters | For example temperature=dropped;max_tokens=renamed:max_completion_tokens; see Parameter compatibility |
X-Prism-Upstream-Endpoint |
A chat request was served through the provider’s Responses API | /v1/responses |
X-Prism-Responses-Translated, X-Prism-Responses-Dropped |
A /v1/responses request was served through Chat Completions |
chat_completions; fields not carried |
X-Prism-Output-Limit-Warning |
A concrete model’s known output limit is below the requested max_tokens |
Warning text |
X-Prism-Budget-Warning, X-Prism-Budget-Active |
A budget threshold applies | Budget state |
X-Prism-Control |
The message was an in-chat /prism command |
accepted or rejected |
X-Prism-Tool-Compression and -Before-Bytes, -After-Bytes, -Saved-Bytes, -Before-Tokens-Estimate, -After-Tokens-Estimate |
Chat requests | Tool-output compression mode and its effect |
X-Prism-Error-Class |
Every error Prism writes | The error code; upstream_error when the provider’s error is passed on |
X-Prism-Upstream-Status |
Provider error passed on | The provider’s HTTP status |
X-Upstream-Request-Id |
Provider sent a request id | The provider’s id |
Retry-After |
overload (5 s), rate limits, a model server that is starting |
Seconds to wait |
Browser pages on another origin can read the headers listed in
Access-Control-Expose-Headers. X-Prism-Adjusted-Params is not in that list.
Pages on other origins can call the main address only when their origin is in
[server] cors_allowed_origins; other origins get 403.
The model itself never sees these headers, and many agent tools drop them. To put the routing result in the response body, see Agent self-description.