By default the workflow host — not the task agents — runs
coderabbit review --agent --type committed against each task branch: once per
dual-review round (alongside the code and expert reviewers) and once per
addendum implementation. Because only committed changes are reviewed, the
ExecPlan durability contract doubles as the review contract. The host parses
the CLI's structured findings; critical and major severities join the
reviewers' blocking items and drive the ordinary fix rounds, while lower
severities are captured without gating integration.
Rate limits are absorbed by the host: a rate-limited review waits a
deterministic 45–90 minutes (coderabbitBackoffMinutes) and retries, up to
coderabbitAttempts total attempts, costing wall-clock but zero agent tokens.
A rate limit that outlives every attempt — or a CLI fault — defers the review
with a documented openIssues entry on the task result instead of blocking
integration; the dual reviewers remain decisive. A CodeRabbit authentication
failure halts the task as fatal-auth.
The run result's coderabbit object reports the effective configuration and
bounded counters (reviews run, findings by severity, rate-limited runs,
deferred reviews). When coderabbitFindingsFile is set, every finding is also
appended as JSONL for cross-run linter tuning.