Deployment info and health

One documented route tells you what a deployment is running before you send it work. Two more exist for operators of a self-hosted deployment: the liveness probe and the cost export. Base URL, authentication and error shapes are described in HTTP API conventions.

GET /deployment

Report the code revision, the model and document bindings, and the MCP memory tools the deployment is serving. It touches no submitted work, so you can call it first to confirm you are talking to the build you expect.

Scope: read.

Parameters

None.

Response

200 with a DeploymentBuildInfo:

{
  "build_revision": "4e1d0c9a7b2f…",
  "model_bindings": {
    "claim_extraction": "…",
    "p1_embedding": "…",
    "fact_adjudication": "…"
  },
  "document_binding_generation": "…",
  "tools": {
    "ingest": 1,
    "pipeline_readiness": 1,
    "delete_document": 1,
    "resolve_entity": 1,
    "claims_and_sources_context": 2,
    "facts_context": 3,
    "combined_context": 4,
    "query_sql": 1,
    "…": 1
  }
}

Field	Meaning
`build_revision`	The source revision stamped into the image when it was built (`REMEMBERSTACK_BUILD_REVISION`). Empty when the image was built without one.
`model_bindings`	The model each pipeline role is bound to, by role name. A self-hosted deployment reports `structure_fallback`, `skeleton_check`, `section_role`, `section_summary`, `claim_extraction`, `relation_normalization`, `entity_resolution`, `fact_adjudication`, `p1_embedding` (the one embedding model, used for every vector), and the OpenRouter settings `openrouter_embedding_provider`, `openrouter_embedding_provider_order`, `openrouter_max_completion_tokens`, `openrouter_reasoning_effort` and `openrouter_reasoning_effort_map`. No secrets.
`document_binding_generation`	The deployment's current document binding generation, or `null`.
`tools`	Each MCP memory tool this deployment serves, by tool name, with its tool version. A tool appears only when the route it calls is served: the seven query tools (`query_sql`, `explain_sql`, `describe_query_space`, `search_query_space`, `list_saved_queries`, `describe_saved_query`, `run_saved_query`) only when the open query space is enabled, and `delete_document` only when document deletion is. A tool's version rises whenever a call it accepts would fail, or mean something else, on a deployment serving the previous version.

The first three fields also appear on every PipelineReadinessReport.

Errors

Status	`detail`	Cause
`403`	`credential may not perform this operation`	The credential has the `ingest` scope, which cannot read.

Example

curl -s "$REMEMBER_API_URL/deployment" \
  -H "Authorization: Bearer $REMEMBER_API_KEY"

from remember import Client
 
memory = Client()
info = memory.deployment_build_info()
print(info.build_revision, info.model_bindings)

GET /healthz

Note

Self-hosted only, and not part of the documented API: it is absent from openapi.json.

Prove the engine process is up and can reach its PostgreSQL database. The Compose file uses it as the API container's health check.

Scope: none. This is the one route that never asks for a credential.

Response

200 with {"status": "ok"} after a SELECT 1 succeeds. If the database cannot be reached, the request fails with 500. While a hard forget runs, it answers 503 with {"detail": {"code": "forget_in_progress"}} like every other route.

Example

curl -s http://localhost:8000/healthz

GET /ops/cost-export/v1

Note

Self-hosted operators only. This route is served by a second, separate listener, not on the API's address, and only when you configure it.

Page through the deployment's cost receipts: one row per model call made by the pipeline workers or by a query route, with tokens, cost and latency. It never carries memory content.

The listener starts only when REMEMBERSTACK_COST_EXPORT_BIND is set (host:port, [ipv6]:port or unix:/path/to.sock). It then requires REMEMBERSTACK_COST_EXPORT_TOKEN, at least 32 bytes long; the process refuses to start without it. See Observability.

Authentication: Authorization: Bearer <REMEMBERSTACK_COST_EXPORT_TOKEN>. This token is separate from the API credential.

Rate limit: one request per second, per process.

Parameters

Name	In	Type	Required	Default	Constraints
`cursor`	query	string	no	start	The `next_cursor` from the previous page.
`limit`	query	integer	no	`100`	1 to 500.

Response

200 with a page:

Field	Type	Meaning
`contract`	string	Always `rememberstack.cost_export.v1`.
`deployment_id`	UUID
`server_time`	date-time	When the page was read.
`horizon`	date-time	The page covers receipts up to this instant.
`cursor`	string	The cursor this page started from.
`next_cursor`	string	Pass it back to continue.
`persist_failures`	integer	Receipts the engine failed to record.
`scope_missing`	integer	Receipts recorded without an attribution scope.
`receipts`	array	The receipts, below.

Each receipt:

Field	Type
`cost_id`	UUID
`deployment_id`	UUID
`source`	`worker` \| `surface`
`work_id`	UUID
`stage`	string or null
`lane`	string or null
`attempt`	integer or null
`surface`	string or null
`call_key`	string
`outcome`	string
`model_name`	string or null
`tokens_in`	integer or null
`tokens_out`	integer or null
`cost_usd`	decimal string or null
`latency_ms`	integer or null
`occurred_at`	date-time (UTC)

Errors

Status	`detail`	Cause
`401`	`unauthorized`	Missing or wrong token.
`422`	`malformed cursor`	The cursor does not parse.
`429`	`rate limited`	More than one request per second.

Example

curl -s "http://127.0.0.1:8001/ops/cost-export/v1?limit=500" \
  -H "Authorization: Bearer $REMEMBERSTACK_COST_EXPORT_TOKEN"