fix: return an iterable stream when interception downgrades stream:true

[likalight]

Fixes #37652.

Root cause

_wrap_response_as_fake_stream (litellm/litellm_core_utils/chat_completion_agentic_loop.py)
returned the bare converted chunk:

return convert_model_response_to_streaming(cast(ModelResponse, response))

convert_model_response_to_streaming returns a single ModelResponseStream —
one chunk, not a stream. Measured on a clean checkout:

returned type : ModelResponseStream
has __aiter__ : False
object field  : chat.completion.chunk

That is both reported symptoms in one line:

1. The crash. Interception downgrades stream: true to a single
   non-streamed call so the agentic loop can run, then hands this value back to
   the streaming path, which does async for over it —
   TypeError: 'async for' requires an object with __aiter__ method, got ModelResponseStream.
   No tool call is needed to reach it; a plain assistant reply is enough.
2. The malformed cache entry. The same object is chat.completion.chunk-shaped
   where a full chat.completion is expected, so a cached copy carries delta
   instead of message and a later read raises KeyError: 'message' inside
   convert_to_model_response_object. The reporter saw this in production
   without an independent repro; the object field above is the mechanism.

The cast at one of the call sites already claimed
"ModelResponse | CustomStreamWrapper", which the returned value never
satisfied.

The fix

Wrap the response in a CustomStreamWrapper over a MockResponseIterator —
the same construction litellm/completion_extras/litellm_responses_transformation/handler.py
already uses to present a non-streamed response as a stream. The result is
genuinely async- and sync-iterable, and the existing cast becomes true.

The helper needs model, custom_llm_provider and logging_obj to build the
wrapper; both call sites already have all three, so they are threaded through as
keyword-only arguments.

This takes option (a) from the issue — make the combination work — rather than
(b) reject it with a 4xx, since the downgrade machinery was clearly built to
support it and only the hand-back was wrong.

Verification

Five tests in tests/test_litellm/litellm_core_utils/test_agentic_loop_fake_stream.py:

• the result exposes __aiter__ and is a CustomStreamWrapper — the reported crash

• it is sync-iterable too

• iterating yields chat.completion.chunk objects that reassemble the original
  content, so the downgrade is transparent to the client

• an already-wrapped stream passes through unchanged (no double wrapping)

• an object with no choices is returned untouched

• tests/test_litellm/litellm_core_utils: 10 failures on this branch and the
  identical 10 on a clean checkout (verified by diffing the failure lists), so
  nothing here regresses; those are pre-existing.

• ruff format --check clean on both touched files; ruff check reports the
  same 5 pre-existing findings on the source file as base, none added.

• Type-discipline checker: identical counts to base on every rule.

Note on the cache symptom

I fixed the shape at its source rather than adding a guard in the cache layer.
If you would also like a defensive check where the entry is written — rejecting
a chat.completion.chunk where a chat.completion is expected — I am happy to
add it, but it seemed better not to paper over the producer inside this fix.

---

On the red code-quality check: it is not from this PR. It fails on every
PR against litellm_internal_staging right now, including ones that touch no
workflow files, because three unit shards in .github/workflows/test-unit.yml
cap the job below the startup-safety invariant. Fixed independently in #38046;
this PR needs no change for it.

[greptile-apps]

Greptile Summary

This PR replaces a bare converted response chunk with a sync- and async-iterable CustomStreamWrapper when interception downgrades a streaming request.

• Threads model, provider, and logging context into the fake-stream helper.
• Preserves existing wrappers and responses without choices.
• Adds coverage for iterable behavior, chunk content, pass-through behavior, and unsupported objects.

Confidence Score: 4/5

The PR appears safe to merge, with only a non-blocking test-organization issue.

The fake-stream implementation restores the expected iterable streaming contract, and no reachable runtime regression was established; the remaining concern is that its tests are separated from the mapped test file.

Files Needing Attention: tests/test_litellm/litellm_core_utils/test_agentic_loop_fake_stream.py

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/chat_completion_agentic_loop.py	Wraps downgraded complete responses with the established stream abstraction; no blocking correctness issue was established.
tests/test_litellm/litellm_core_utils/test_agentic_loop_fake_stream.py	Adds focused regression coverage, but places it in a new module rather than the repository’s existing mapped test file.

_{Reviews (1): Last reviewed commit: "style: apply ruff format to the new agen..." | Re-trigger Greptile}

[greptile-apps]

Regression tests use separate module

This bug fix places all five regression cases in a new module instead of extending the existing mapped test file for chat_completion_agentic_loop.py, fragmenting coverage and making the cases easier to omit from targeted test runs.

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codecov]

Codecov Report

❌ Patch coverage is 80.00000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...litellm_core_utils/chat_completion_agentic_loop.py	80.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing likalight:fix/code-interpreter-fake-stream-iterable (77e9281) with litellm_internal_staging (d447be1)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (f005afa) during the generation of this report, so d447be1 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩