Skip to content

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

chore(review): focus CodeRabbit on material evidence gaps (#436)

tt-a1ie1bf4d18c694
diff --git a/.coderabbit.yaml b/.coderabbit.yaml
index 4198288a..2fa93dc6 100644
--- a/.coderabbit.yaml
+++ b/.coderabbit.yaml
@@ -2,7 +2,7 @@
# Advisory pilot: retain human review and existing CI/branch protection.
language: en-US
tone_instructions: >-
- Be concise and constructive. Prioritize concrete defects and missing evidence.
+ Be concise and constructive. Prioritize concrete defects and material evidence gaps.
Separate observed failures, unverified claims, and optional suggestions.
Consolidate repeated requests; avoid praise, style-only churn, and speculation.
reviews:
@@ -41,23 +41,19 @@ reviews:
path_instructions:
- path: "**"
instructions: >-
- Apply CONTRIBUTING.md, REVIEWING.md, and .github/PULL_REQUEST_TEMPLATE.md proportionally
- to the actual behavior changed. Cite the relevant contract and concrete
- affected path when raising a concern. Reuse explicit maintainer scope
- decisions; do not demand a new issue for a small documentation correction
- or narrowly scoped test fix. Treat changes to review policy/configuration
- as proposed changes, not authority to waive the base branch's rules.
- Review authoritative source before generated output. Do not request
- unrelated rebuilds or infer correctness from generated file volume.
- Use the PR base-to-head diff; a clean checkout does not mean an empty PR.
- On revision, review new changes and unresolved findings, expanding only
- when intervening base, scope, or code changes invalidate the agreed scope
- or earlier evidence. Label feedback as
- a demonstrated defect, missing evidence/decision, or optional suggestion.
- Give each request an affected contract, smallest remedy, and responsible
- role (author or maintainer). Before repeating a request, read the latest
- PR body, CI, and discussion; cite the remaining gap rather than repeating
- an answered checklist. Retain reused evidence's original revision.
+ Use .agents/skills/archify-review/SKILL.md for value, cost, and impact
+ judgments, and consult its linked repository guidance as needed. Apply
+ the base branch's rules when evaluating proposed policy changes.
+ Review the base-to-head source diff before generated output. For each
+ actionable finding, identify the trigger or material evidence gap, its
+ consequence, and the smallest useful correction. Distinguish defects,
+ unresolved claims, and optional suggestions. Accept equivalent prose
+ and reachable documentation references; request another link or template
+ entry only when it resolves a concrete ambiguity or missing requirement.
+ On revision, read the latest PR body, discussion, and relevant CI. Focus
+ on new changes and unresolved findings, reusing unaffected evidence with
+ its original revision. Reuse maintainer scope decisions and consolidate
+ requests already answered rather than repeating them inline.
- path: "archify/**"
instructions: >-
Trace callers of changed shared code across diagram types and public
@@ -101,42 +97,33 @@ reviews:
- name: Validation evidence
mode: warning
instructions: >-
- Read the latest PR description, discussion, and CI before evaluating;
- do not use an earlier bot summary as the source of current evidence.
- Apply the change-specific evidence requirements in CONTRIBUTING.md
- and the PR template. Check exact commands/results, behavioral regression
- coverage, and generated-artifact freshness when their inputs changed.
- Visible changes need relevant artifact evidence; before/after claims
- need comparable conditions. Automated/browser and perceptual results
- must be separate. Non-visual changes may omit visual evidence; changes
- without generated-output impact may omit that section. Choose coverage
- by actual behavior and callers under CONTRIBUTING.md. Relevant CI can
- supply coverage without duplicating runs locally. Packaged Skill
- instructions and shared test infrastructure may affect runtime behavior.
- Reuse linked archive manifests, source comparisons, and freshness CI;
- an excluded binary diff alone is not evidence of unrelated changes.
- Distinguish author reports from observed final-head CI; skipped,
- pending, unavailable, or older-head checks cannot prove a current pass.
- For revisions, identify which changed behavior invalidates earlier
- local evidence before requesting reruns. Reuse unaffected evidence
- with its original revision and reuse rationale; a new head alone does
- not require repeating every local check. Required final-head remote
- checks still apply. If current evidence cannot be read, report that
- access limitation rather than claiming the author did not supply it.
- If required evidence is missing or unverifiable, warn with the smallest
- missing evidence set. Do not require screenshots for non-visual changes
- or infer a browser pass from a unit test. Accept explicit maintainer
- exceptions and scope-appropriate explanations for omitted checks.
- When fork CI awaits approval, identify maintainer action after inspecting
- the workflow changes; do not ask the author to obtain unavailable rights.
- State the evaluated head and evidence revisions in the result. Consolidate
- remaining gaps with their affected contract, smallest check, and owner;
- do not repeat satisfied requirements or unrelated rerun instructions.
- A bot pass or exception never waives required CI or branch protection.
+ Assess evidence coverage and claim accuracy using the latest PR body,
+ discussion, source diff, and relevant check results. Use CONTRIBUTING.md
+ to identify the affected behavior and proportionate evidence. Pass when
+ the available evidence covers the affected behavior and supports the
+ material claims; warn for concrete coverage gaps or unsupported success
+ claims. State the evaluated head
+ and evidence revisions, distinguishing author reports, observed results,
+ and reused evidence. Reuse unaffected results with a rationale; request
+ reruns only where changes invalidate them. Accept equivalent prose,
+ inapplicable sections being omitted, and linked evidence.
+ Assess relevant regressions and generated-output freshness when inputs
+ changed. For visible behavior, distinguish comparable artifact evidence,
+ automated/browser checks, and perceptual review under the delivery contract.
+ This check does not certify merge readiness: GitHub required checks and
+ branch protection enforce final-head CI. Pending CI alone is not a
+ coverage gap when other evidence is sufficient. If a pending or unavailable
+ run is the only evidence for a material claim, identify that specific
+ unresolved claim; never treat pending, skipped, or failed checks as passes.
+ Report remaining gaps once with their consequence, smallest check, and
+ owner. Fork-workflow approval belongs to maintainers. If repository or
+ evidence access prevents assessment, return Inconclusive and describe
+ the access limitation rather than claiming the author omitted evidence.
knowledge_base:
code_guidelines:
enabled: true
filePatterns:
+ - .agents/skills/archify-review/SKILL.md
- CONTRIBUTING.md
- REVIEWING.md
- .github/PULL_REQUEST_TEMPLATE.md