lsm / neokai / 25271366292

82%

push

github

Retry SDK API requests on transient connection errors (#1762)

* Retry SDK API requests on transient connection errors

- Detect fetch-level errors (socket connection was closed, verbose: true, etc.)
  and categorize them as ErrorCategory.CONNECTION instead of SYSTEM
- Auto-retry the query once on transient connection errors that escape SDK
  retries, following the same pattern as startup timeout retries
- Sanitize raw fetch error strings before surfacing to the user: never show
  "verbose: true" or developer-facing fetch internals in the UI
- Exclude transient connection blips from circuit breaker error counting so
  a single network drop doesn't trigger the breaker unnecessarily

Co-Authored-By: Claude Opus 4.7 <[EMAIL]>

* Narrow transient error patterns and replace weak tests with integration tests

- Narrow `fetch failed` to `TypeError: fetch failed` and remove broad `connection closed` pattern from both query-runner and circuit-breaker
- Replace weak boolean-duplicating transient error tests with integration tests that exercise QueryRunner.start() and verify spy behavior (retry count, sanitized messages, CONNECTION category)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* Reorder fatal error check before transient filter in circuit breaker

Check fatal error patterns (extractErrorPattern) before skipping transient
substrings, so messages containing both a fatal "Connection error" inside a
stderr block and a transient "connection reset" substring still increment
circuit-breaker counts. Previously the transient check ran first and could
silently mask persistent outages.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* Use String(error) instead of error.message for transient error detection

For native TypeError instances (e.g. fetch failures), error.message is just
"fetch failed" without the "TypeError:" prefix.  Using String(error) ensures
the full error name+message is available for pattern matching, so the
existing 'TypeE... (continued)

7851 of 10832 branches covered (72.48%)

Branch coverage included in aggregate %.

93 of 93 new or added lines in 3 files covered. (100.0%)

62349 of 87214 relevant lines covered (71.49%)

253.59 hits per line