fix(tools): coerce read tool offset/limit to int — kills agent on string-typed args#110
Open
triscacezar-droid wants to merge 1 commit intohuggingface:mainfrom
Open
Conversation
The local read handler was doing `max(args.get("offset") or 1, 1)` which
TypeErrors on Python 3 if the model passes "1" (string) instead of 1 (int):
File "agent/tools/local_tools.py", line 146, in _read_handler
offset = max((args.get("offset") or 1), 1)
TypeError: '>' not supported between instances of 'int' and 'str'
This kills the agent loop mid-task. Some models (and intermediate JSON
serialization steps) occasionally hand back numeric tool args as strings,
even though the tool schema declares them as int. The same risk applies
to `limit`.
Fix: coerce both via int() with a typed-fallback to the default, then
clamp offset to >= 1 as before. No behavior change when args arrive as
ints (the common case).
Hit while running a long agent investigation; the agent had completed
~30 minutes of useful work, then died on a mid-run file read.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Symptom
The local `read` tool handler crashes the agent loop with:
```
File "agent/tools/local_tools.py", line 146, in _read_handler
offset = max((args.get("offset") or 1), 1)
TypeError: '>' not supported between instances of 'int' and 'str'
```
This happens whenever the model returns `offset` (or, by extension, `limit`) as a string like `"1"` instead of an integer. Python 3's `max()` can't compare across str and int.
Why this happens
Tool args go through JSON serialization on the way out of the model and on the way back into the handler. Even though the tool schema declares `offset: integer`, some models occasionally emit it as a string (especially smaller / faster ones, or when the model is mid-conversation and the response shape drifts), and LiteLLM passes it through verbatim. Every other tool handler that takes numeric args has the same latent risk.
Impact
The crash kills the entire agent run mid-task. In my case it ate ~30 minutes of an investigative ml-intern run that had just completed a useful baseline phase — losing the live state.
Fix
Coerce `offset` and `limit` via `int()` with a typed-fallback to the existing default. No behavior change when args arrive as ints (the common path).
```python
try:
offset = int(args.get("offset") or 1)
except (TypeError, ValueError):
offset = 1
try:
limit = int(args.get("limit") or DEFAULT_READ_LINES)
except (TypeError, ValueError):
limit = DEFAULT_READ_LINES
offset = max(offset, 1)
```
Compatibility
Pure additive defensiveness — passes all existing call shapes. No new dependencies. The same pattern could reasonably be applied to other numeric tool args (`bash.timeout`, etc.) in a follow-up; happy to do that here if you'd prefer one bigger PR.