A rate-limited reviewer filled a task's Chat tab with 14 identical
"Reviewer using model: ..." markers and no review text, hammering an
already-limited provider.
Root cause: the reviewer was the only AI lane that never classified
provider errors, so a 429 became an UNAVAILABLE verdict. With no
validator fallback configured the fallback ladder re-ran the SAME model
instantly, and fn_review_step answered with "code review remains
blocking; retry once" — bounding the loop with prompt text rather than
code. The tool's catch-all also swallowed the error into tool output, so
withRateLimitRetry, UsageLimitPauser and RetryStormError never fired.
- reviewer: throw ReviewerProviderError for usage-limit/transient errors
instead of laundering them into UNAVAILABLE, and never spend the
fallback budget (which bounds bad reviews) on an outage.
- reviewer: absorb flaky-network blips in-lane via withRetry with
jittered backoff; rate limits still escalate immediately.
- executor: re-raise the fatal after the prompt via
throwDeferredReviewerFatal — pi-agent-core converts tool throws into
tool_error results, so a tool cannot throw out of session.prompt().
- executor: give code review a real MAX_CODE_REVIEW_UNAVAILABLE_RETRIES
counter, mirroring the plan/spec limiter.
- reviewer: dedupe the model marker on text, so same-model retries stay
silent while a genuine model switch still emits.
Also fixes the run-on rendering: AgentLogType gains `status` for complete
engine messages. `text` means "streamed delta" and is re-glued with
join(""), which is why N standalone markers rendered as one string. The
split is at the type, not a separator — a separator would reintroduce the
FN-5787/5789/5803 streamed-spacing regression.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
