OpenAI subscription stream failures are fatal after the first model event

[danielkov]

Problem

An OpenAI subscription response stream can terminate the ACP session after model output has started, even when the underlying failure is a transient response-body interruption. Kit retries the same failure before the first model event, but Kit 0.1.86 treated it as fatal after an event had been emitted.

Version

kit 0.1.86

Environment

• OS: Darwin 25.5.0
• Architecture: arm64
• Surface/host: ACP
• Provider: openai-subscription

Reproduction

1. Start an ACP session using openai-subscription.
2. Begin a model turn and allow at least one ModelTurnEvent to be emitted.
3. Interrupt the HTTP response body before the provider sends response.completed (for a deterministic test, serve an SSE response with a truncated Content-Length).

The reported session ran for approximately 54 seconds after its last completed tool result before the response stream ended.

Actual behavior

The turn terminates with a fatal provider error and does not persist a completed assistant item:

openai-subscription stream transport failed
(timeout=false, connect=false, request=false, body=false, decode=true)

Sanitized fatal diagnostics:

{
  "stage": "stream",
  "retryable": false,
  "attempt": 1,
  "reqwest": {
    "timeout": false,
    "connect": false,
    "request": false,
    "body": false,
    "decode": true
  },
  "source_chain": [
    { "kind": "hyper" },
    { "kind": "io", "classification": "unexpected_eof" }
  ]
}

Reqwest reports decode=true because Response::bytes_stream() wraps response-body frame failures as decode errors. The low-level cause was unexpected_eof.

Expected behavior

Kit should recover from a transient post-event interruption without assuming that a non-deterministic model will replay the same SSE frames. After ACP v2 is available, Kit can replace the in-progress assistant message with the new authoritative response. Recovery must not duplicate text or reasoning, and tool calls must not execute twice.

Dependencies

• Blocked by feat(acp): add v2 mid-turn steering #24, which moves Kit's ACP integration to v2. The recovery design depends on ACP v2 message identity and replacement semantics; raw provider-frame replay is not a safe basis for recovery.

Related work

• Draft implementation: fix: replace interrupted OpenAI responses safely #10
• Earlier pre-event stream retry work: 8729c84

[walker-tx]

Observed another occurrence on a newer Kit version.

Environment

• Kit: 0.1.91
• OS: Darwin 25.6.0
• Architecture: arm64
• Surface: ACP
• Provider: openai-subscription

Behavior

An ordinary agent turn terminated with the existing fatal stream-transport error rather than completing:

openai-subscription stream transport failed
(timeout=false, connect=false, request=false, body=false, decode=true)

The sanitized schema-v2 diagnostics reported stage stream, retryable: false, attempt 1, reqwest decode: true with timeout/connect/request/body all false, followed by a Hyper source and an I/O source classified as unexpected_eof. The source chain was neither unknown nor truncated.

This confirms the post-event decode=true / unexpected_eof failure remains reproducible in 0.1.91. Kit marked it non-retryable on the first attempt.