fix(kennel): read the agent name from RunContext.agent, not four dead probes

[breedx]

The bug

_agent_name_from_context probes four attributes, none of which exist:

for attr in ("agent_name", "name"):      # not on RunContext
    ...
for attr in ("agent_name", "name"):      # nor on deps
    ...
return "unknown"                          # so: always this

Against pydantic-ai 2.35:

RunContext.__dataclass_fields__ has "agent_name"  -> False
RunContext.__dataclass_fields__ has "name"        -> False
RunContext.__dataclass_fields__ has "agent"       -> True

The agent is on .agent, the name on Agent.name. So no probe can match, and the "best-effort fallback" is the only reachable branch — every call, not an edge case.

agent_wing() then faithfully renders agent:unknown.

Why it matters

The agent wing is documented as one agent's own cross-project reflections. In practice it is one shared bucket for every agent in the process — not private, not per-agent. Anything written there trusting the docstring is mis-scoped.

Observed in a deployment: kennel_list_wings returned only agent:unknown (2 drawers) and repo:<path> (218). No per-agent wing had ever existed.

What made it hard to spot: repo-wing bylines are correct. They come from the agent_run_end callback parameter (register_callbacks.py:42 → record_run_end → metadata["agent"]), not from this probe. So the name is plainly available in the process; only this lookup misses it.

The fix

Read ctx.agent.name. One field, no probe.

RunContext.agent is typed Agent | None and an unnamed Agent is legal, so "no name" is a real answer — returned as "" rather than a stand-in. _resolve_wing raises KennelScopeError at the one point something actually asks to be scoped by the agent.

Refusing at the lookup instead would break callers that never needed the name — kennel_remember defaults to the repo wing — so the refusal sits where the mis-scoped write would happen. (I tried it the other way first; it broke 8 existing tests, all of them repo-wing writes.)

Tests

tests/test_puppy_kennel_agent_wing.py, 7 cases. Three fail against the current implementation, including "two agents do not share one wing" — both collapse to agent:unknown today.

They use SimpleNamespace, not Mock, deliberately: a Mock auto-vivifies every attribute, so it satisfies the old probe and hides which field is actually read. That property is plausibly why this shipped.

• kennel suite: 58 passed
• full suite: 2122 passed, 2 skipped
• ruff check + ruff format --check clean

Note

This does not migrate existing agent:unknown drawers. After the fix they simply stop being written to; the old bucket remains until someone decides what to do with it. Happy to follow up if you'd like a migration.