Skip to content

Turn any idea, plan, or codebase into a beautiful interactive diagram. An agent skill for Claude Code, Codex, and more.

7.9 KB
Raw Blame

CodeRabbit review cost and latency controls

Checked official documentation on 2026-10-10. This note separates documented mechanisms from hypotheses; no configuration, review command, or external setting was changed by this research.

Practical finding

The narrow candidate is to exclude only the 30 checked-in HTML copies corresponding to the 15 golden outputs, while retaining their generators, runtime template, schema, authored JSON, and other HTML. This reduces duplicate review input, but any latency benefit remains unmeasured. Filtering also changes checkout and coverage; successful generated-output CI establishes freshness, not visual quality or CodeRabbit review of the excluded files.

Documented controls and limits

Control Documented mechanism Performance and coverage implication
reviews.path_filters Review include/exclude patterns also affect Git sparse checkout. Less input is plausible; excluded files may also be absent from tool context. Configuration reference
Generated-file handling Defaults ignore generated directories/suffixes, build output, lock files, archives and minified assets; they do not blanket-ignore HTML. Exclusions win regardless of order; ordinary positive includes restrict scope. Review globs and sparse-checkout matching differ. Excluded files disappear from walkthrough/Change Stack, though GitHub retains its full file list. Use exact negative paths alone for these copies. A source retained by review matching does not establish the exact sandbox inventory. The vendor describes filtering as faster, without a quantitative promise. Path controls and ignored paths
Incremental review auto_incremental_review: true is the default: each eligible push reviews new commits since the previous review. false leaves subsequent pushes for manual review. Drafts default to skipped. auto_pause_after_reviewed_commits defaults to 5; 0 disables the pause, and the counter resets when lifted. Reduce redundant triggers during rapid iteration. This controls review frequency, not the duration of one review. Label opt-in with automatic reviews disabled still permits later incremental reviews while the label remains. Automatic review controls
Manual pause/trigger PR comments @coderabbitai pause and resume stop/restart automatic review. review is incremental; full review reviews the whole PR from scratch. @coderabbitai ignore belongs in the PR description. Batch ready changes before another pass. Pause is documented as stopping automatic reviews; it is not a documented cancellation API for an already-running review. Review commands
Cache disable_cache defaults to false; setting true refetches code/dependencies each run. The FAQ explicitly connects encrypted code/dependency caching with faster reviews. Keep the default. No per-repository speedup magnitude is documented. Configuration reference, FAQ
Profile and summaries Current profiles are quiet/chill/assertive; chill is default. Quiet groups some feedback into the summary. Summary, changed-file summary and sequence-diagram switches govern generated content; a summary placeholder still triggers generation with high_level_summary: false. No documented latency guarantee. Profiles also change tool strictness where fallback configuration is used. Fewer visible comments do not prove less analysis. Configuration reference, Tool behavior
Pre-merge checks Built-in/custom checks perform separate agentic validations; off disables a check and warning/error change enforcement. Docstrings can return indeterminate when eligible files cannot be analyzed. Turning these off removes evidence. The vendor's “faster reviews” claim concerns earlier feedback to humans, not measured bot runtime reduction. Built-in checks
Queue/timing No user-set debounce interval, scheduling delay, queue priority, concurrency cap or review-runtime timeout was found in the reviewed public reference/commands. Rate-limit availability is provider-controlled; the FAQ says rapid pushes can supersede reviews while still consuming allowance. Scope and trigger controls are available; provider queue time remains a confounder. Absence here is a documentation finding, not proof that no private capability exists. Configuration reference, Review commands, FAQ

Tools run in sandboxes and are chosen based on repository files. Losing an excluded output can change what tools can inspect or what an invoked command can load. A golden comparison can protect exact reproduction from retained inputs, but cannot replace independent review of those inputs or inspect a deployed artifact's behavior. Third-party tools

Repository evidence and trial

The source assessment is pinned to 0c5550a6e5c7c3fc42ab6f2ed2c96d223461dbe3: golden inventory contains 15 outputs with copies under examples/ and archify/examples/; CI scope classification independently selects generated checks for output-only changes. In PR #800, those 30 paths account for 5,550 of 9,221 changed lines across 69 files. That is potential input reduction for one PR, not a runtime estimate.

Observed on 2026-10-10: PR #810 pending 05:27:02 → completed 05:30:57 UTC (235 seconds); PR #808 final pending 02:16:54 → completed 02:29:33 UTC (759 seconds). These include queue and processing and are not a comparable before/after experiment.

A small trial can reuse normal PR work:

  1. Record head/base SHA, effective configuration, full changed-file/line counts, excluded counts, reviewed commit range, trigger type and push count. Keep profile, checks, cache and summary settings fixed while testing the exact filters; separate first/full reviews from incremental ones.
  2. Capture the CodeRabbit-specific pending → terminal transition for each head and attempt. Read the full commit-status history, using each record's created_at, rather than only combined status. If the current progress surface is a check run, retain its ID, timestamps, status and conclusion via check runs. Missing timestamps are unavailable evidence. Do not change the reporting surface just to measure it.
  3. Compare several similar scopes, avoiding pushes during a measured attempt; record superseded/error/rate-limited attempts separately. Report durations and sample sizes, not a promised percentage gain. Without internal telemetry, label the interval end-to-end wait.
  4. Inspect reviewed/ignored paths and tool errors, and retain generated-check evidence at the final revision. Excluded, skipped or indeterminate files/checks are coverage gaps, never passes. Retain exact filters only if their source boundary remains correct; assess speed separately from freshness and visual acceptance.

Start with the narrow filters and batching existing review triggers. Profile changes, summary removal and disabled checks would mix coverage changes into the first latency trial and are not justified by these sources alone.