cometchat-audit by @cometchat

Ground truth: the authority for "correct" is RULES.md + the resolved family's skills (cometchat-<family>-core/-production/-troubleshooting) and the offline probe npx @cometchat/skills detect --json. Verify every symbol/prop you flag against the family catalog + docs (via -core/references/docs-map.md); never flag from memory. This is a read-only review — produce the report first; only change code if the user says so (then hand each fix to the owning skill).

Use this skill when

Someone wants an existing CometChat integration checked — before a launch, a security review, an upgrade, or a "why is this flaky / is this safe" question. It is the proactive complement to the per-family troubleshooting skills (cometchat-<family>-troubleshooting, which react to a specific symptom).

How to run the audit

Detect — npx @cometchat/skills detect --json: framework, installed UI Kit, version_conflict, existing_cometchat. Resolve <family> from peers.yaml. If nothing is detected, say so and stop (nothing to audit).
Read the integration — the init/login/lifecycle file, where the surface renders, the env/secret files, the server token endpoint (if any), and any custom message/theme/feature code. Do NOT read node_modules/.d.ts.
Score each check below, gather evidence (file:line), and write a ranked report — most severe first — with, per finding, the fix and the skill that owns it. Then offer to apply fixes; don't auto-edit.

The checklist

Security (highest severity)

Auth Key present in client code / bundle / binary? → critical; must move to server-minted tokens (cometchat-security, cometchat-<family>-production).
Token endpoint deriving the UID from a client parameter (?uid=)? → impersonation.
REST API Key anywhere client-side or in the repo? → critical.
Sessions revocable on offboarding (flush tokens)? RBAC roles configured, or everyone on default? (cometchat-security.)

Correctness

init() → login() → render order honored; no component rendered before login resolves; no init during SSR; StrictMode/double-invoke guarded.
initFromSettings used (ai-agent attribution), not the classic builder init.
Surface has real dimensions (not a collapsed 0-height box); host CSS not leaking into .cometchat.
No raw localization keys rendered (group_info etc.).
No dead affordances — every default-on control wired or hidden.
Listeners removed on unmount/dispose (no duplicate messages/leaks).

Version & drift

version_conflict from detect (UI Kit major ≠ the family's target)? → route to cometchat-<family>-migration.
Kit/SDK versions pinned (not a floating major)? iOS: the exact kit+SDK+Calls pins.

Production-readiness

HTTPS everywhere; logout teardown (session + push tokens); dashboard extensions/AI enabled on the PRODUCTION app; error handling in place (React families: an error boundary such as CometChatErrorBoundary; iOS/Android/Flutter: SDK error listeners / kit error-state views). (cometchat-<family>-production.)

Accessibility & localization (enterprise procurement)

Focus rings / contrast not stripped by global CSS; chat container labelled; reduced-motion honored (cometchat-a11y).
Locale set before render; no raw keys; RTL handled if targeted (cometchat-i18n).

Compliance (if in scope)

App in the right data-residency region; a data-deletion path exists (cometchat-compliance); moderation enabled if the product needs T&S (cometchat-moderation).

The report format

Lead with a one-line verdict (ship / fix-first / not-ready) and counts by severity. Then a table: Severity · Finding · Evidence (file:line) · Fix · Owning skill. List what PASSED too, so the reader knows it was checked. End with the top 3 things to fix first. Do not inflate severity, and do not flag a "problem" you didn't verify against docs/catalog.

Common pitfalls (of the auditor)

Flagging from memory — verify each symbol/prop against the catalog + docs before calling it wrong.
Auto-editing — this is read-only; report, then ask.
Reading node_modules/.d.ts — audit the user's code + docs, not kit internals.
Severity inflation — a dev-only Auth Key in a local .env is not the same as one shipped in production; judge by what actually reaches users.

Verify it works

The report names the framework + family, lists ranked findings each with file:line + a fix + an owning skill, states what passed, and gives a clear verdict — and nothing was edited unless the user approved.