mirror of
https://github.com/block/buzz.git
synced 2026-11-10 11:21:05 +00:00
85996fdff4
## Summary
Split out of the `flakey-monday` flake-fix series. Three Rust unit tests
in `buzz-cli`, `buzz-agent` and `buzz-admin` depended on wall-clock
timing or on a production-sized pool budget and failed on loaded CI
runners. Each now waits on, or counts, the thing it actually asserts.
Commits (cherry-picked from `flakey-monday`, original messages kept):
- `test(cli): throttle the first desktop sign-in by count, not clock`
- `test(agent): wait for both responses in
steer_rejected_on_empty_prompt`
- `test(agent): pin steer_rejected_on_empty_prompt to the empty-prompt
guard`
- `test(admin): give the NIP-FI Postgres tests a test-lane pool budget`
### Root cause and fix per test
**`client::desktop_throttle_tests::a_throttled_sign_in_reconnects_with_fresh_evidence`**
(`crates/buzz-cli/src/client.rs`): failed once in CI with `conns` left
1, right 2.
- Root cause (confirmed): the fake desktop throttled a sign-in only when
`last.elapsed() < 300ms`, and `last` was captured before the provider
handshake and credential setup. When that setup took over 300 ms, the
first sign-in was signed straight away, nothing was throttled, the
client never reconnected, and the relay saw one connection.
- Fix: the fake desktop throttles exactly its first `sign_event` call
via an atomic counter, which is the scenario the test describes. The
relay's unsigned-socket deadline goes from 200 ms to 5 s, as in the
sibling deadline tests, so it no longer shapes the scenario. Assertions
are unchanged: two connections, only "c1" signed, a fresh assertion on
the reconnect.
**`steer_rejected_on_empty_prompt`**
(`crates/buzz-agent/tests/fake_llm.rs`): panicked with "empty steer
prompt was not rejected".
- Root cause (confirmed): `session/prompt` runs in a spawned task while
`_goose/unstable/session/steer` is answered inline from the stdin loop.
The fake LLM finishes the turn almost at once, so the prompt response
can arrive before the agent has read the steer. The receive loop broke
on the prompt response and never saw the steer's -32602.
- Fix: keep reading, bounded at 40 messages as before, until both the
steer and prompt responses have arrived in either order. The follow-up
commit also asserts `error.message == "steer: prompt must not be
empty"`, because `steer_session` returns -32602 for "no active run to
steer" and an `expectedRunId` mismatch too, which are exactly the
replies a steer gets when the turn ends first. The empty-prompt check
runs right after decoding, so the message is deterministic in both
orderings.
**`nip_fi::postgres_tests::{backfill_binds_only_bindable_hosts_and_reports_the_rest,
backfill_skips_a_community_being_deleted,
bind_refuses_hosts_outside_the_suffix_and_unbind_keeps_the_last_issuer}`**
(`crates/buzz-admin/src/nip_fi.rs`): `connect to test DB:
Sqlx(PoolTimedOut)` after about 3.5 s.
- Root cause (confirmed): the tests built their pool from
`DbConfig::default()`, sized for a warm relay pod (3 s acquire timeout,
two eager spare connections). In the first wave of the PostgreSQL lane
every nextest worker clones its database from the template at once, so a
first connection can take longer than 3 s.
- Fix: keep the production writer pool (`Db::new`, so the session policy
is still exercised) but set `min_connections` 0 and a 30 s acquire
timeout, the budget sqlx's `PgPool::connect` defaults give the other
PostgreSQL tests. `Db::connect_cold_start` uses the same values.
### Related issue
N/A (flake fixes; none found).
### Testing
Human testing is still required. This PR does not carry
`buzz-review-completed`; the author has not yet run these tests
themselves on this branch.
#### Verification
These changes were verified locally on the `flakey-monday` branch before
the split:
- buzz-cli throttle test: old code under CPU saturation (40 busy loops
on 32 cores) failed every iteration with `left: 1`, the CI failure. New
code, `desktop_throttle_tests` with `--test-threads=16`: 30/30 passed
unloaded, 15/15 under the same saturation.
- buzz-agent steer test: forced CI ordering (temporary 500 ms delay
before sending the steer) made the old test panic and the new test pass.
`fake_llm steer` tests with `--test-threads=16`: 30/30 passed. The
follow-up ran 100/100, and was falsified: with the empty-prompt guard
removed and the steer sent after the turn, the code check still passed
and the new message assertion failed.
- buzz-admin NIP-FI tests, against docker Postgres: the three tests, 10
iterations at `--test-threads=3`: 10/10. Behind a proxy delaying each
new connection by 4 s: the old config failed with the CI `PoolTimedOut`
at 3 s, the new config passed.
Note: `crates/buzz-cli/src/client.rs` also changed on `main` since the
series forked. The cherry-pick applied cleanly; only hunk offsets differ
from the original commit.
Nothing was re-run on this split branch: the workstation was overloaded,
so no tests, builds or linters were executed after cherry-picking. CI on
this PR is the verification for the split branch.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Signed-off-by: mtoohey <mtoohey@squareup.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>