azmaveth / ex_mcp / bde822a85c937c0d7ad736eae05dbe5c9eef5fa6

65%

Build Type

push

github

Committed by azmaveth

Commit Message

test(client): make 'transport closure in ready' test telemetry-driven

Replaces a fixed-sleep + state-poll pattern that flaked under CI load
(`:disconnected` instead of `:ready` at the assertion point) with a
telemetry-subscription pattern that waits for the actual state
transition events the state machine already emits.

## What changed

1. **`Arbor.LLM.Client.StateMachine` emits `pid: self()` in the
   `[:ex_mcp, :client, :state_transition]` telemetry metadata** so
   async tests with multiple state machines can filter events to
   their own client. Other handlers ignore the extra key.

2. **`TestTransport.respond_to_handshake/2` polls internally** for
   the handshake request to arrive at the test transport. Previously
   it failed immediately with `{:error, :no_messages}` if called
   between the `:enter :handshaking` callback and the asynchronous
   `:start_handshake_internal` message that actually sends the
   request. New default 2s window (configurable) covers CPU-busy CI
   runners.

3. **The failing test subscribes to state-transition telemetry**
   (`attach_state_transition_listener/1` helper at the bottom of the
   module) and uses `assert_receive` instead of `Process.sleep` +
   `StateMachine.get_state`. The handler filters by `metadata[:pid]`
   so async tests don't see each other's transitions. Auto-detaches
   on test exit.

4. **The final assertion matches `from_state: :ready`** to pick the
   transition triggered BY the transport closure, not earlier
   transitions that `assert_receive` left in the mailbox during the
   handshake flow.

## Why this commit only touches one test

The same `Process.sleep(N) + assert state.state == X` pattern shows
up 25 more times in this file. The pattern is timing-sensitive and
is the right shape to migrate eventually. This commit establishes
the pattern (helper, telemetry pid, polling respond_to_handshake)
on the one test that actually failed CI — leaving the rest as a
focused follow-up rather than r... (continued)

Coverage Stats

0 of 9 new or added lines in 1 file covered. (0.0%)

1 existing line in 1 file now uncovered.

1434 of 16001 relevant lines covered (8.96%)

5.85 hits per line