provider-openai: recovered incomplete streaming response can leave unpaired function_call items, permanently breaking the session

[danshapiro]

Summary

provider-openai has a recovery path that salvages non-completed streaming responses:

[PROVIDER] OpenAI recovered a non-completed streaming response (status=incomplete...): returning the partial response instead of failing.

(The warning references amplifier-support#339; that repo is not publicly resolvable, so this issue is filed here.)

When the salvaged partial response contains a function_call item, the salvage can leave that call unpaired (no function_call_output) in persisted conversation history. Because the Responses API requires every function_call item in replayed input to have a matching function_call_output, every subsequent request in the session fails — the session is permanently wedged and no retry can ever succeed.

Observed behavior

• Model: gpt-5.6-sol, with enable_response_chaining: auto.

• A chained continuation response (llm:response event with continuation_count: 1, output_tokens: 2050, cost_usd: null, 2026-08-01) ended non-completed and was salvaged by the recovery path.

• The salvaged partial contained:

  1. A function call with truncated/empty arguments {}, which failed tool schema validation ("Command is required").
  2. A function_call item (call_NpD59KEhgybRvh2IyJ4AEWKF) that never received a function_call_output.

• Every subsequent request in that session failed with:

  InvalidRequestError: "No tool output found for function call call_NpD59KEhgybRvh2IyJ4AEWKF"

  14 occurrences in the session log, including retries 7 hours later. Because the dangling function_call is persisted in history, retries can never succeed.

Note: the caps were NOT the cause

The request budget was max_output_tokens=128000 with a 900k context, and the turn was at ~110k input / 2k output. So the non-completed status was likely content_filter or a stream drop — but the warning text only guesses "max_output_tokens".

Proposed fix

When salvaging a non-completed response (especially mid-continuation):

1. Drop function_call items whose arguments are truncated/unparseable.
2. Guarantee pairing: every retained function_call must get an output — synthesize an error function_call_output (e.g. "call truncated by provider") for any salvaged call that tooling didn't execute.
3. Log incomplete_details.reason explicitly in the salvage warning instead of guessing at the cause.

Affected code

amplifier-module-provider-openai — the streaming recovery/salvage path and response chaining.