What changed

Vector DB

• A text over 2,000 characters is refused with 400 input_too_long (param names it), also when sent with values. Metadata keys that start with _hk_ are refused, and metadata plus text over 10,240 bytes is 400 metadata_too_large.
• A namespace written to must be 1–64 letters, digits and . _ - : @ + = (an id or an email fits). Writes into an older namespace with a space or a slash are refused with 400 invalid_request; reading from it still works.
• Filters with $or, $and, $exists or an unknown operator are refused with 400 filter_operator_unsupported, and filter errors carry code and param.
• Visitor chat searches only collections opened to it (visitorChat: true). Every existing collection starts closed. A visitor's chat run searches by text across the whole collection, every namespace included, with at most 10 matches: open only a collection whose every vector may be shown to anyone.
• A collection's endpoint is its own address (/api/vectors/:id), and it has visitor_chat. Deleting vectors by id no longer lowers vector_count at once; GET /api/vectors/:id reads the current count.
• The connector's query_vectors no longer takes returnMetadata and caps topK at 50.

AI Memory

• A store created without kind is kind "memory" (it used to be "kv"), and every data call on it needs userId (400 user_id_required). Stores created before keep their kind and work as before.
• Store answers (GET/POST /api/projects/:id/memory, GET /api/memory/:id, /api/resources/memory) no longer carry provider, endpoint or seq; they gain searchable, edge_fallback, user_id_required, created_at and quota.
• Changing a memory's text gives it a new id: use the id in the answer (previous_id is the old one). 409 already_exists when the user already has that text.
• Refusals carry code and param: a bad user id is 400 invalid_user_id, a call the store's kind does not have is 409 wrong_store_kind. HTTP statuses are unchanged.
• The connector's search_memory sends scope as given; with userId too and the two different, it is 400 conflicting_param.

AI Gateway

• An unknown model is 404 model_not_found (it was 400), on /api/ai/chat and /api/ai/v1.
• A model you name that cannot be reached is 503 model_unavailable, with nothing charged, instead of being answered by Llama. Use harakumo-1 to fail over.
• /api/ai/v1/chat/completions refuses unknown fields and the parameters it does not honour with 400 unsupported_parameter; parameters with no effect are accepted only at their no-effect values. The native POST /api/ai/chat refuses tools, tool_choice and response_format other than text (other unknown fields are still ignored).
• Output defaults to each model's own: 4,096 tokens on vendor models, 1,024 on Llama, up to 32,000. A max_tokens over a model's most is a 400 naming the model.
• Tools and JSON output (json_object, json_schema) work only on the models whose supported_parameters list them, once the model has passed a live check through the gateway; elsewhere they are 400 unsupported_parameter. json_object needs the word JSON in a message, and a conversation with tool calls must send tools.
• Every cheap and mid-price model is on by default for a workspace that never saved its model settings; Claude Opus stays off, and so does Kimi K2.6 until it is confirmed.
• Added, nothing removed: OpenAI error objects carry param and metadata, usage has cost, and harakumo has request_id.

Agents

• Agents and runs have one shape: tools is a list (it may hold function tools), allow_writes is true or false, trigger is { kind, label }, a run has cost_usd only, stop_reason is null when it completed, started_by is api when someone asked, and a step has tool, parsed arguments and ok true or false.
• GET /api/agents/:id answers recent_runs (the 5 newest, without steps) instead of runs (20 with steps). GET /api/agents/runs/:runId has one run with its steps, and GET /api/agents/:id/runs pages through them.
• A run you ask for starts at once in the background and gets its id at once (POST /api/agents/:id/runs); wait (0 to 60 seconds) counts from when the request arrives. A run that finished answers 200 (it was 202). The queued answer is gone.
• maxSteps counts model calls, each with up to 8 tool calls. A run has 5 minutes (timeLimitSeconds 10 to 300), and a cost cap of $0.25 on Hobby and $1 on paid plans unless the agent or the run sets one (stop_reason budget).
• A run with no sign of life for 5 minutes is marked failed (stale) and frees its slot. A run that could not start is 503 run_not_started, nothing done or charged.
• A schedule, app events and visitor chat can be on together: setting one no longer turns the others off, and trigger.kind is a label only.
• An agent whose tools are function tools only is offered only those; list tool names beside them to keep them.
• SDK: agents.run resolves to the run itself (not { run }), and waits for it to end.

Old names kept until 2026-11-23

userId on vector calls (the same field as namespace), memory for a memory's text, and user_id or scope for userId are accepted names, not old ones: they stay.