mirror of
https://github.com/headroomlabs-ai/headroom.git
synced 2026-11-22 13:29:19 +08:00
fix(backends/litellm): tolerate a bare string in tool_result content list (#3795)
## Description
`_convert_messages_for_litellm` flattens a list-shaped `tool_result`
content like this:
```python
if isinstance(tr_content, list):
tr_content = "\n".join(
b.get("text", "") for b in tr_content if b.get("type") == "text"
)
```
This assumes every element of the list is a dict. But a client may put a
**bare string** in the list (`content: ["result text"]` — a valid
Anthropic tool_result shape). `b.get(...)` on a `str` raises
`AttributeError: 'str' object has no attribute 'get'`, which 500s the
entire request via `--backend litellm` / bedrock.
Confirmed:
```python
_convert_messages_for_litellm([{"role":"user","content":[
{"type":"tool_result","tool_use_id":"tu_1","content":["a bare string result"]}]}])
# -> AttributeError: 'str' object has no attribute 'get'
```
Fix: accept bare strings (contribute their text), keep
`{"type":"text",...}` dict blocks, and skip other block types — the same
guard the sibling any-llm converter uses. Plain-string content and
text-block lists are unchanged.
(The neighboring missing-`tool_use_id` / missing-`id` `KeyError`s are
intentionally left as-is: defaulting `tool_call_id` to `""` would
produce an unroutable tool message and mask the error with a confusing
downstream failure, so a hard failure there is preferable to a
silent-wrong request. This PR is scoped to the unambiguous bare-string
crash.)
## Type of Change
- [x] Bug fix (non-breaking change that fixes an issue)
- [ ] New feature (non-breaking change that adds functionality)
- [ ] Breaking change (fix or feature that would cause existing
functionality to change)
- [ ] Documentation update
- [ ] Performance improvement
- [ ] Code refactoring (no functional changes)
## Changes Made
- `headroom/backends/litellm.py`: the `tool_result` list-content flatten
now handles bare-string elements and skips non-text blocks, using a
dedicated `text_pieces` local (avoids colliding with the function's
existing `list[dict]` `parts`).
- `tests/test_backend_bugs.py`: added
`test_tool_result_list_content_with_bare_string_block` — a list mixing a
bare string, a text block, and an image block flattens to the two text
pieces without raising.
## Testing
- [x] Unit tests pass (`pytest`)
- [x] Linting passes (`ruff check .`)
- [x] Type checking passes (`mypy headroom`)
- [x] New tests added for new functionality
### Test Output
```text
tests/test_backend_bugs.py -k ConvertMessagesToolBlocks -> 8 passed
uvx ruff@0.16.8 check headroom/backends/litellm.py tests/test_backend_bugs.py -> All checks passed!
uvx mypy@1.20.2 headroom/backends/litellm.py -> clean (no error in this file)
```
## Real Behavior Proof
- Environment: Windows 11, Python 3.12.11, project venv, pytest 9.1.1,
ruff 0.16.8 and mypy 1.20.2 via uvx.
- Exact command / steps: called `_convert_messages_for_litellm` with a
`tool_result` whose content list contains a bare string — before the fix
it raised `AttributeError`; after, the tool message content is the
joined text. Verified plain-string content, text-block lists, and mixed
bare-string+text+image lists all produce the right joined text with no
crash. Red/green: the new test fails (AttributeError) with the fix
reverted; all 8 tool-block tests pass with it.
- Observed result: a tool continuation whose result content is a list
containing a bare string no longer 500s the request.
- Not tested: a live litellm/bedrock round-trip (the message conversion
is covered by unit tests).
## Runtime Rollout Safety
- Rollout-managed feature(s): none.
- Minimum rollout channel: N/A.
- Stable/default behavior changed: no for conformant inputs (plain
string, text-block lists convert identically); a previously-crashing
bare-string list now converts instead of 500ing.
- Kill switch / disable path: N/A.
- Unsafe override required: no.
- Qualification impact: none.
- Rollback path: revert this commit; a bare-string tool_result content
element again raises AttributeError.
## Review Readiness
- [x] I have performed a self-review
- [x] This PR is ready for human review
## Checklist
- [x] My code follows the project's style guidelines
- [x] I have performed a self-review of my code
- [x] I have commented my code, particularly in hard-to-understand areas
- [ ] I have made corresponding changes to the documentation (N/A:
internal message conversion)
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective or that my
feature works
- [x] New and existing unit tests pass locally with my changes
- [x] I did **not** edit `CHANGELOG.md`
## Additional Notes
Mirrors the `isinstance(b, dict)` guard used by the any-llm backend's
tool_result handling, so both backends tolerate the same client input
shape.
Co-authored-by: JD Davis <mxjerrett@gmail.com>
This commit is contained in:
@@ -997,9 +997,18 @@ class LiteLLMBackend(Backend):
|
||||
for tr in tool_result_blocks:
|
||||
tr_content = tr.get("content", "")
|
||||
if isinstance(tr_content, list):
|
||||
tr_content = "\n".join(
|
||||
b.get("text", "") for b in tr_content if b.get("type") == "text"
|
||||
)
|
||||
# A tool_result content list is usually
|
||||
# ``{"type":"text",...}`` blocks, but a client may put
|
||||
# a bare string in the list. ``b.get`` on a str raised
|
||||
# AttributeError and 500'd the whole request; accept
|
||||
# bare strings and skip non-text/other blocks.
|
||||
text_pieces: list[str] = []
|
||||
for b in tr_content:
|
||||
if isinstance(b, str):
|
||||
text_pieces.append(b)
|
||||
elif isinstance(b, dict) and b.get("type") == "text":
|
||||
text_pieces.append(b.get("text", ""))
|
||||
tr_content = "\n".join(text_pieces)
|
||||
tool_msg: dict[str, Any] = {
|
||||
"role": "tool",
|
||||
"tool_call_id": tr["tool_use_id"],
|
||||
|
||||
@@ -274,6 +274,33 @@ class TestConvertMessagesToolBlocks:
|
||||
assert converted[0]["role"] == "tool"
|
||||
assert converted[0]["content"] == "Line 1\nLine 2"
|
||||
|
||||
def test_tool_result_list_content_with_bare_string_block(self):
|
||||
"""A tool_result content list may contain a bare string, not just
|
||||
``{"type":"text",...}`` blocks. ``b.get`` on a str raised AttributeError
|
||||
and 500'd the whole request; bare strings must be accepted and other
|
||||
block types skipped."""
|
||||
backend = self._make_backend()
|
||||
messages = [
|
||||
{
|
||||
"role": "user",
|
||||
"content": [
|
||||
{
|
||||
"type": "tool_result",
|
||||
"tool_use_id": "toolu_03",
|
||||
"content": [
|
||||
"bare string result",
|
||||
{"type": "text", "text": "typed block"},
|
||||
{"type": "image", "source": {"type": "base64", "data": "x"}},
|
||||
],
|
||||
},
|
||||
],
|
||||
},
|
||||
]
|
||||
converted = backend._convert_messages_for_litellm(messages)
|
||||
assert converted[0]["role"] == "tool"
|
||||
# bare string + text block joined; the image block is skipped.
|
||||
assert converted[0]["content"] == "bare string result\ntyped block"
|
||||
|
||||
def test_assistant_tool_use_with_text(self):
|
||||
"""Assistant message with both text and tool_use blocks."""
|
||||
backend = self._make_backend()
|
||||
|
||||
Reference in New Issue
Block a user