vstorm-co / subagents-pydantic-ai / 31982653888

100%

push

github

fix(toolset): name the exception's class in a task-status answer (#81)

`check_task` and `wait_tasks` composed their failure lines straight from
`handle.error`:

```python
f"Error: {handle.error}"                              # check_task, FAILED
f"Retry {handle.retry_count}: {handle.error}"         # check_task, RETRYING
f"Outcome: {handle.error}"                            # check_task, CANCELLED
f"{handle.status.value.upper()} - {handle.error}"     # wait_tasks
```

`handle.error` embeds the exception's own message —
`f"{type(exc).__name__}: {exc}"`,
`f"Transient error (retry {n}): {exc}"` — and a model client's message
carries the
failing request URL with the key still in its query string when the
profile has a
custom `base_url`.

## Why this one is worse than the frame it looks like

What a status tool returns is a tool **return**. A host stores a
`ToolReturnPart`
whole — deliberately, because a return is the tool's own answer rather
than
something the framework composed — and renders it to everyone who can
read the run.
So this is the surface a host's scrubber cannot reach: it can strip a
`RetryPromptPart`, and several do, but not this.

The `RETRYING` line also leaks for delegations that **eventually
succeed**. `finish`
clears `error` on completion, but by then a model that polled mid-retry
already has
the answer on a stored row.

## The fix

One helper, `error_for_model`, reading `TaskHandle.exception` — the
field 0.2.19
added for exactly this — and all four lines go through it. The class is
named; the
message is not. `handle.error` is untouched and still carries the whole
thing, which
is the contract #80 established: a host reads the class from
`exception`, composes
its own sentence, and logs the original.

The `usage limit exceeded` marker survives. It is this library's own
text, not a
provider's, and it is the thing that tells a model to stop delegating
rather than
retry — so `usage limit exceeded: <exc text>` becomes
`usage limit exceeded:... (continued)

11 of 11 new or added lines in 2 files covered. (100.0%)

1605 of 1605 relevant lines covered (100.0%)

1.0 hits per line