vstorm-co / subagents-pydantic-ai / 32568343849

100%

push

github

fix(prompts): send a return description the way pydantic-ai sends one (#84)

Each of the ten delegation tools carried its own `Returns:` line in the
middle of its prose. pydantic-ai wraps a docstring's `Returns:` section
in `<summary>` and `<returns>` tags instead — what every tool built from
a docstring already sends — so a host registering these beside tools of
its own ends up with two conventions in one tool list, and the
structural half of `task`, `wait_tasks` and `check_task` was a paragraph
the model had to notice rather than a field it was handed.

Every tool text is a `ToolText` (summary / usage / returns) now and
renders into the framework's shape. The test pins that shape against a
tool pydantic-ai renders itself rather than against a string literal, so
a change on that side fails here rather than leaving this library
speaking the old dialect.

**Four of the ten said nothing about what came back at all.** They do
now, and it is the part worth having — `check_task` returns a result in
full where `wait_tasks` may cut one and mark it, and re-delegating a
task to "finish" a marked result is exactly the mistake that sentence
prevents.

The dynamic half — the subagent list on `task`, the model list on
`delegate` and `create_agent` — renders **inside** the summary. Appended
after it, as a first attempt did, the tags bracket half the text; there
is a test for that too.

`delegate` also stops restating its own parameters. `## Parameters`
listed five arguments the docstring already describes, which is where
the model actually reads them, and nothing compared that copy against
the signature.

Verified: 1197 tests pass, coverage 100%, ruff and pyright clean.
Released as 0.2.21.

38 of 38 new or added lines in 1 file covered. (100.0%)

1635 of 1635 relevant lines covered (100.0%)

1.0 hits per line