lancekrogers / claude-code-go / 42323876671

83%

push

github

fix: wait only before actual retries (#32)

Rate-limit failures currently wait after the final allowed attempt,
delaying the exhausted-retries error. Earlier failures also incur
retry-after plus a separate backoff wait.

Wait once before each actual retry, using the larger of jittered backoff
and the preceding rate-limit delay. Server-dominated waits receive up to
20% positive jitter above the server minimum, so concurrent callers
retain retry spread. Duration conversion and jitter addition saturate
instead of overflowing.

Return exhaustion immediately after the final failure. If the parent
context cancels a failed CLI attempt, return its context error even on
the final attempt. Preserve retry counts and exhausted-error wrapping.
Document delay and timeout semantics.

Tests cover exhaustion with and without hints, a single wait,
rate-limit-to-network transitions, cancellation during the final running
subprocess and retry waits, nonretryable errors, convenience wrappers,
jitter bounds/variability, and duration overflow. Scheduling tests use
an internal per-call wait dependency to record delays without wall-clock
thresholds; subprocess cancellation tests exercise real contexts. The
original exhaustion/additive-wait tests and the follow-up
cancellation/jitter regressions were verified failing before their
respective fixes.

Validation:
- `just test quick`
- `just test race`
- `just lint`
- `just build examples`
- Local subprocess probes verify immediate exhaustion, one rate-limit
wait, and `errors.Is(err, context.DeadlineExceeded)` after cancellation
of the final CLI attempt. No real service calls.

Work item: **WI-62ce60** — `workflow/design/ccgo-issue-31-retry-delay`
in `lancekrogers/claude-code-go-planning`. The branch worktree is linked
to this work item and commits carry its traceability tag.

Fixes #31

38 of 38 new or added lines in 2 files covered. (100.0%)

2300 of 2776 relevant lines covered (82.85%)

0.95 hits per line