openai-python 3.15/3.16 and openai-node 7.18/7.19: Cache Prewarming, Webhook Management, connector_id Deprecation

The official OpenAI SDKs shipped six releases in a single day, 2026-09-18: openai-python v3.15.0 / v3.16.0 / v3.16.1 / v3.16.2 and openai-node v7.18.0 / v7.19.0. Where the previous batch (3.14.x / 7.16–7.17) was themed on reliability, this one is themed on features: a warming switch for the prompt cache, a management REST surface for webhook endpoints, a formal deprecation of the MCP tool's connector_id, and a session layer for WebSocket connections. This article walks each item back to its PR and type sources (all verified on 2026-09-19 via the GitHub API against release notes, PR diffs, and the v3.16.2 tag); example code comes from the official README and type definitions.

1. Release timeline

The rhythm is worth noting: python and node shipped nearly in pairs this time — the four features (agent session settings, audio-mini, WS sessions, prewarm) landed in python v3.15.0 and node v7.18.0, while webhook management and the connector_id deprecation landed in python v3.16.0 and node v7.19.0. The two packages still do not share a version scheme; the same feature landing in different versions per language is the norm.

2. Prompt-cache prewarming: prewarm

prompt_cache_options gains prewarm (boolean, default false). The official docstring reads: "Prepares the prompt cache without generating output. Defaults to false. When set to true, overrides the generate field to false."

The problem it solves: before a long session issues its real request, warm the stable prefixes — system prompt, tool definitions — into the cache so the real request hits it. The canonical pattern is the WebSocket warmup flow: send a generate: false response.create event to prepare state without model output, then send the real request — the official README calls this warmup; prewarm: true lives in prompt_cache_options, and per its docstring setting it true forces the generate field to false.

response = client.responses.create(
    model="gpt-5.6-terra",
    input=[
        # long, stable system prompt and tool definitions
    ],
    prompt_cache_options={
        "mode": "implicit",
        "ttl": "30m",
        "prewarm": True,
    },
)

Two existing constraints are unchanged: prompt_cache_options requires gpt-5.6 and later models, and ttl still only accepts 30m. The diagnostics surface (prompt_cache_diagnostics) can verify whether your post-warmup request actually hit the cache (covered in our earlier dedicated article).

3. Webhook endpoint management: client.webhooks

A new top-level resource, client.webhooks, with a full endpoint lifecycle:

A subresource, client.webhooks.event_types.list(), fetches the available event types. The subscription enum has 18 values in seven families: batch.* (completed / failed / expired / cancelled), response.* (completed / failed / cancelled / incomplete), eval.run.* (succeeded / failed / canceled), fine_tuning.job.* (succeeded / failed / cancelled), realtime.call.incoming, video.* (completed / failed), safety.alert.created.

endpoint = client.webhooks.create(
    event_types=["response.completed", "safety.alert.created"],
    name="prod-responses",
    url="https://example.com/hooks/openai",
)
save_secret_somewhere_safe(endpoint.signing_secret)  # plaintext exactly once

One rule to remember about the key model: the signing_secret comes back in plaintext only at creation and rotation; every other read exposes only a masked signing_secret_hint. Also note the official comment on updated_at: tests and unchanged updates do not advance it — which makes it a reliable signal for whether an endpoint's configuration truly changed.

4. MCP connector_id deprecation: two migration paths

The connector_id field on MCP tools is now marked deprecated: true in the OpenAPI spec, with the official docstring reading: "This field is deprecated for models released after September 1, 2026. Use server_url to connect to a remote MCP server, or tunnel_id to connect through a Secure MCP Tunnel."

Three qualifiers to parse: first, what is deprecated is the connector_id style of service-connector direct connection, not MCP tools themselves; second, the cutoff is per model — models released after 2026-09-01, while models released before that date are outside the deprecation statement; third, there are two migration paths, server_url (your own remote MCP server) or tunnel_id (a Secure MCP Tunnel).

The Responses, beta, and Realtime MCP tool types (Mcp, RealtimeResponseCreateMcpTool, and others) are all annotated. If your code still uses connector_id against service connectors such as Dropbox or Gmail, and your model list may move to post-09-01 versions, migrate to server_url or tunnel_id before upgrading. The protocol-level MCP knowledge is unaffected — our MCP channel explainers still apply.

5. Managed Responses WebSocket sessions: lane routing and final-response collection

The old problem with a raw client.responses.connect(): when several responses run concurrently on one connection, events interleave, routing is hand-rolled stream_id bookkeeping, and finalization means guessing which event is terminal. This batch adds a session layer in both languages (officially, managed sessions) — same idea, different API names.

On python it is the openai.lib.responses_websocket module (requires the openai[realtime] extra):

from openai import AsyncOpenAI
from openai.lib.responses_websocket import AsyncResponsesWebSocketSession

async with AsyncOpenAI() as client:
    async with client.responses.connect() as connection:
        async with AsyncResponsesWebSocketSession(connection) as session:
            lane = session.lane("conversation")          # register a named lane
            await lane.send({                             # stream_id added automatically
                "type": "response.create",
                "model": "gpt-5.6-luna",
                "input": "Say hello.",
            })
            response = await lane.get_final_response()   # drains remaining events, returns terminal Response

(Adapted from the official README, with the model name swapped to this site's current family; the README also demonstrates six budgets via ResponsesWebSocketLimits(max_lanes=8, max_events_per_lane=128, ...).)

Key semantics, from the official README: lane.send adds the stream_id to a response.create and rejects conflicting routing metadata; get_final_response() consumes remaining events and returns the terminal Response (including failed / incomplete), and a repeated call returns the cached result until a newer response.created appears on that lane; unrouted, unknown, and connection-level events go to the default lane (session.default) — watch it when using named lanes; budget overflow raises ResponsesWebSocketBufferError and closes the owned connection (drop the connection rather than silently drop events); lane budgets persist until a physical reconnect. The README also documents warmup (generate=False), tool turns, forks, and compaction over WS.

On node it is ResponsesWebSocketSession from openai/lib/responses/responses-websocket-session: lane.create(request) to send, lane.receive({ signal }) for raw events, lane.finalResponse({ signal, maxResponseBytes }) for the terminal response; budgets are maxLanes / maxBufferedEvents / maxBufferedBytes, exhaustion fails and drains the largest backlog (bytes first, then event count), and nested API error events raise a WebSocketError carrying the original event.

6. The rest in brief

• Agent session model settings (python #3882 / node #2755): client.beta.agents.sessions.update(id, agent={...}), hitting POST /v1/agents/sessions/{id}. Three fields — model (string), reasoning.effort (enum none / minimal / low / medium / high / xhigh / max), and service_tier (auto / default / flex / priority / fast). The docstring's scoping deserves emphasis: "Model settings for subsequent turns. Omitted fields stay unchanged."
• Responses accepts audio models (python #3886 / node #2759): the Responses model union adds gpt-audio-mini and gpt-audio-mini-2025-12-15; on the ChatModel enum, gpt-5.1-mini moves to the legacy tail. Note both audio models belong to the legacy audio family on this site's deprecation calendar (migration window 2027-01-20) — new choices do not change their lifecycle status.
• python v3.15.0 catches up on compaction progress events (#3866): ResponseCompactionCompactingEvent (beta variant with an optional agent.agent_name), fields type / item_id / output_index / sequence_number, fully aligned with node v7.17.0 — closing the cross-package gap from our previous breakdown.
• Two python fix releases: v3.16.1 avoids loading unrelated API resources on first use (#3898, shortening the startup path); v3.16.2 drops TextFormatT parameterization in parse_response to fix a memory leak (#3084 / #3088) — long-lived processes should go straight to v3.16.2.
• Assorted fixes: python v3.15.0 preserves chat stream moderation results (#3864), clarifies incoming SIP call ID usage (#3885), and updates image request examples (#3889); node v7.18.0 validates WebSocket results and preserves header defaults (#2763).

7. Upgrade notes and common pitfalls

• Going straight from v3.14.x to v3.16.2, or from v7.16/v7.17 to v7.19.0, is fine — the two batches do not conflict.
• prewarm not taking effect: confirm the model is gpt-5.6 or later (a precondition of prompt_cache_options as a whole), confirm you are not also sending a generate: true WS event (prewarm overrides it), then verify hits with prompt_cache_diagnostics.
• Lost webhook signing_secret: there is no second plaintext read; rotate_secret resets it (match keep_old_secret_active_for_24_hours against your rollout window).
• connector_id calls failing or flagged: check your model against the deprecation cutoff (models released after 2026-09-01) and move to server_url or tunnel_id.
• WS session events that belong to no lane: that is the default lane's job (unrouted, unknown, connection-level); consume session.default when using named lanes.
• Faster startup after upgrade, but resource attribute errors: v3.16.1 made resources lazy; audit code that relied on implicit eager loading of client.resources (the official fix reads: avoid loading unrelated API resources on first use).

8. Next steps

• The Responses API Compaction Progress Event: response.compaction.compacting and compaction_trigger — the full breakdown of the event whose cross-package gap this article closes in section 6.
• openai-python 3.14.x and openai-node 7.16/7.17: Stream Error Normalization, WebSocket Backpressure, and SSE Fixes — the previous reliability batch; this batch's WS session layer builds on its backpressure work.
• openai-python 3.9/3.10 and openai-node 7.11/7.12: Prompt Cache Diagnostics, API Key Expiry, GPT Image 2.5 — the backstory of prompt_cache_diagnostics and prompt_cache_options, which you will want for verifying prewarm hits.
• gpt-6-astra appears in the OpenAI SDK: the new ChatModel ID and the Safety Alerts API — the system behind the safety.alert.created entry in the webhook event enum.
• Codex CLI 0.154.0 and SDK 0.154.0: Worktrees, ExternalMessage, and the ultra Reasoning Effort — the Codex-side September release explainer; see our latest article for CLI 0.155.x.