fix(lm): handle None usage from Responses API on truncated responses

[isaacbmiller]

Problem

OpenAI's Responses API returns usage=None for incomplete/truncated responses (e.g. when max_output_tokens is hit or content filtering triggers). LiteLLM passes this through as-is across all supported versions (1.64.0–1.82.6).

The existing code dict(getattr(response, "usage", {})) only guards against the attribute being absent — not against it being explicitly None. Since getattr returns the attribute value when it exists (even if None), this becomes dict(None) and raises TypeError: 'NoneType' object is not iterable.

Fix

Add or {} to coerce None to an empty dict at all 3 call sites:

• dspy/clients/base_lm.py:109 — history entry construction
• dspy/clients/lm.py:197 — sync forward() usage tracking
• dspy/clients/lm.py:238 — async aforward() usage tracking

Tests

Added two new tests (test_responses_api_with_none_usage and test_responses_api_with_none_usage_async) that mock a truncated Responses API response with usage=None and verify the sync and async paths handle it gracefully.

[greptile-apps]

Greptile Summary

This PR fixes a TypeError crash when OpenAI's Responses API returns usage=None for truncated responses (e.g. max_output_tokens exceeded or content filtering). It adds or {} at the three call sites where getattr(response, "usage", {}) is used, so an explicit None value is safely coerced to an empty dict. The PR also bundles a secondary fix in alitellm_completion: the async streaming path now correctly applies _add_dspy_identifier_to_headers before calling _get_stream_completion_fn, matching the existing sync path's behavior.

Confidence Score: 5/5

Safe to merge — targeted fix with no logic regressions and solid test coverage for both sync and async paths.

All three None-usage sites are fixed consistently. Tests cover the sync and async paths end-to-end with realistic mocked truncated responses. The secondary alitellm_completion header fix brings async streaming in line with the sync path. No P0/P1 findings remain.

No files require special attention.

Important Files Changed

Filename	Overview
dspy/clients/base_lm.py	Single-line fix: adds or {} to coerce None usage to an empty dict in the history entry, matching the fix in lm.py.
dspy/clients/lm.py	Applies the or {} guard at two usage-tracker call sites (sync and async forward), and separately fixes async streaming to pass the DSPy identifier headers — bringing it in line with the existing sync path.
tests/clients/test_lm.py	Adds two targeted tests (sync and async) that mock a truncated Responses API response with usage=None and assert the fix behaves correctly end-to-end.

Sequence Diagram

sequenceDiagram
    participant Caller
    participant LM
    participant LiteLLM
    participant OpenAI Responses API

    Caller->>LM: forward() / aforward()
    LM->>LiteLLM: litellm.responses / litellm.aresponses
    LiteLLM->>OpenAI Responses API: POST /responses
    OpenAI Responses API-->>LiteLLM: response (usage=None on truncation)
    LiteLLM-->>LM: ResponsesAPIResponse(usage=None)
    Note over LM: getattr(response, "usage", {}) → None<br/>None or {} → {}<br/>dict({}) → {}  ✅
    LM->>LM: history entry: usage = {}
    LM->>LM: usage_tracker.add_usage(model, {})
    LM-->>Caller: results

Loading

_{Reviews (1): Last reviewed commit: "fix(lm): handle None usage from Response..." | Re-trigger Greptile}