fix: Parlant tool descriptions drift from SDK master models

[AlexanderZ-Band]

Summary

• Parlant hand-wrote its 12 tool docstrings instead of reading the master Pydantic models in runtime/tools.py, so a master reword never reached Parlant agents — it had already drifted (stale unprefixed add_participant(...)/remove_participant(...) example calls for tools actually named band_add_participant/band_remove_participant).
• Adds a band_tool decorator that sets each function's __doc__ from get_tool_description(func.__name__) before registering with p.tool — the tool name is never retyped as a string (it's just func.__name__, always written to match its TOOL_MODELS entry), and decorator syntax is preserved.
• Only genuine Parlant-specific addition kept: a mentions-format note on band_send_message (Parlant takes a comma-separated string; the master model describes a list).
• Adds a regression test (test_descriptions_match_master_source) asserting every Parlant tool description starts with its master get_tool_description() text.

Fixes INT-1170.

Test plan

• uv run --extra dev-parlant pytest tests/integrations/parlant/test_tools.py -v --no-cov — 39 passed (real Parlant SDK, not mocked)
• uv run ruff check / uv run ruff format --check
• uv run pyrefly check

🤖 Generated with Claude Code

https://claude.ai/code/session_0114NCg97KJ6mNNK5RmvsbpH

[linear-code]

INT-1170

[amit-gazal-band]

Reviewed the diff and ran it live: booted examples/parlant/01_basic_agent.py against a real Band platform with parlant 3.3.2 and drove it as a user through three turns (greet + band_get_participants; band_lookup_peers + band_add_participant; membership re-check). It works — the agent replies, completes multi-tool turns, and the add-participant really landed. tests/integrations/parlant/: 44 passed on this head.

Most of what I found on the previous head (4e32a5b2) — every parameter reaching the LLM with description=None, the lost @john/agent-name handle format on band_add_contact, the lost identifier guidance on band_add_participant, the lost either/or hints on the contact tools — is fixed by 4a0b5794. I re-dumped every constructed ToolEntry on this head and confirmed each one now carries real per-argument text. Nice fix, and the Annotated[...]-wrapping approach keeps the single-source-of-truth property intact.

One gap survives it, inline below. Everything else from my earlier pass is resolved.

One thing outside this diff, for a follow-up: examples/parlant/01_basic_agent.py:89 and :110 instruct the agent to call band_add_participant / band_remove_participant "with the name parameter" — both take identifier. In my live run the model dropped the argument once ([ToolCaller] Inference::Completion: Argument 'identifier' is missing) before recovering, so it costs a wasted inference round rather than breaking the turn. The restored identifier description makes it much less likely; the guidelines are still worth correcting.

[AlexanderZ-Band]

Addressed both points from the review, pushed in two commits:

examples/parlant/01_basic_agent.py:89/:110 (the "outside this diff" follow-up): fixed — both guidelines now say identifier instead of name, matching AddParticipantInput/RemoveParticipantInput's actual field. (a38c9b3)

sent_status inline comment: replied inline with the fix + why a bare Literal doesn't work here (parlant's schema builder only special-cases real enum.Enum, and crashes on Literal[...] — verified live). Also deduplicated the master's own Literal["pending", "approved", "rejected", "cancelled", "all"], which was independently hand-typed in two places (ListContactRequestsInput and ListSentContactRequestsInput) — one ContactRequestSentStatus alias now, matching the existing EventMessageType pattern. (048ac46)

tests/integrations/parlant/: 46 passed (44 + 1 new regression test) on dev-parlant; tests/runtime/, tests/core/, and the rest of the default-venv suite (4499 passed) unaffected. ruff check/format and pyrefly check clean on the touched files.