Editorial prompt runtime status
GET /prompts/{name}/runtime explains which editorial source ordinary chat is configured
to use and what its serving services last resolved. The only managed name is
agent-editorial. Publishing an immutable version or moving a Langfuse label does not
enable runtime management. Playground selects its candidate independently.
The route requires the configured API key and a verified admin principal. JWT deployments
ignore X-User-* headers and require a verified admin bearer token. Missing credentials
return 401, verified non-admins 403, and authentication disabled 503. The temporary
API-key-only GET /prompts exception never grants runtime-status access. Unknown names
return 404. Successful responses send Cache-Control: no-store.
Response
| Field | Meaning |
|---|---|
name, environment, deployed_ref | Managed name and this API process's configuration. The ref can be null. |
required_label, configured_label | Environment-required label and actual configuration. With management off they can differ. |
tracing_enabled, prompt_management_enabled | Independent integration and runtime-selection flags. |
configured_source | file unless both flags are enabled; otherwise langfuse. This is configuration, not a fetch result. |
remote_available | Whether the administrative read succeeded. Remote failure returns 200 with false so serving observations remain readable. |
remote_selection | Null when unavailable or versionless; otherwise version, label and selection_reason. |
paths | One entry each for /query and /query/v2: route enabled, service initialized, and last_resolved. JSON and SSE share the corresponding service. |
The remote selection_reason is environment_label when the configured environment label
selects the version; otherwise newest_version. The latter can happen with configured
latest or a missing environment label. The legacy /prompts response keeps its existing
active_version and label fields; these describe administration, not runtime adoption.
last_resolved is null until that serving service successfully resolves an artifact. When
present it contains:
{
"resolved_at": "2026-11-11T10:00:00Z",
"prompt": {
"name": "agent-editorial",
"version": null,
"label": "production",
"source": "file",
"contract_version": "agent-editorial-v1"
},
"content_sha256": "<SHA-256 of canonical consumed messages>"
}
This example is a file resolution, even though its configured label is production.
Sources are file, langfuse, langfuse_last_good, or an explicitly injected test
candidate. Ordinary Playground requests never populate either normal path's observation.
Restart/service replacement clears observations. Reading status never constructs or
resolves a provider. No prompt body, credentials, question, passage or user identifier is
included.
The hash covers the ordered {role, content} message objects encoded as UTF-8 JSON with
ensure_ascii=false and compact separators. The per-answer audit field
editorial_prompt contains this hash and the same provenance; the stored answer's
synthesis.prompt retains its five provenance fields.
Display and rollback
Show remote label selection separately from configured source and last observed artifact. Label an unobserved service accordingly. A resolution can precede publication or fail later; concurrent requests can publish in a different order. This status is not a language compliance verdict or an approval record. Confirm each answer's stored provenance and trace when verifying rollout or rollback.
The SDK can return stale cached content with source=langfuse while refreshing, including
after repeated refresh failures. Same-process admin moves invalidate its cache, but external
moves and in-flight responses can leave stale content. A publish response or elapsed TTL
does not prove rollback completion. Production activation and label moves retain their
independent review, evaluation and change-record requirements. Code-composed refusal/error
notices keep their existing localization.