feat(teams): show that an agent is working — eyes on the message, card in its thread (CHOO-2315)

[amaudruz]

Teams gets the acknowledgement Slack already has.

• 👀 on the message that asked, for the length of the turn — added when the agent starts, removed when it goes idle. Plain Bot Connector call (PUT/DELETE /v3/conversations/{conversation}/activities/{activity}/reactions/1f440_eyes) on the same regional host and app-only token the bridge already uses to send, so no Graph permission and no delegated user. It renders in a channel, a group chat and a 1:1 alike — the only signal that does.
• The status card lands in that message's thread rather than at the channel root (runtime_state_follows_anchor), because Teams renders a thread inline under its root post.

Which message gets marked is recovered on the way in: a runtime-state report names the thread, not the message inside it, so inbound messages record what asked, per thread and per conversation. Inside a thread that is a reply, not the root; a chat has no thread at all.

The honest negatives

Written down in the bridge README rather than faked:

• No per-agent @handle. A Teams app manifest admits one bot (bots is maxItems: 1) and a tag — the nearest analogue to a Slack user group — takes user ids only, so a bot cannot be in one. The exception is an Entra agent user account, which is @-mentionable like a person but preview-gated and costs a directory object and a licence per agent.
• Native streaming not used. The streaminfo "informative update" is documented as one-on-one chats only, one request per second, with a hard two-minute cap — a response-latency surface, not a signal for work that runs for minutes.
• Typing indicator not relied on. Transient, undocumented as to scope, and in practice absent in channels.
• The reaction endpoint has no Bot Connector REST reference page — Microsoft documents the capability through the Teams SDK. A tenant whose Teams service has not shipped it answers an error; the mark is skipped with a warning and the turn carries on with the card alone.

Tests

test_teams_adapter.py — the mark goes on the message that asked and not the thread root, comes off when the turn ends, is added once per turn, survives awaiting-input, covers every message an agent marked, works in a chat with no thread, and leaves the card intact when the tenant refuses the reaction. test_teams_clients.py — the reaction URL shape, empty body, id escaping, and error propagation.

Not yet verified against a live tenant: which conversation types actually render the reaction. Worth a check before this is announced.

🤖 Generated with Claude Code

[abeldantas]

Resolved the merge conflicts with main (branch predated the 0.23.0 release cut and the Teams README/adapter rewrites)