Dispatching any subagent whose task involves writing or editing WordPress PHP — the compliance check must happen in the controller, before the first subagent is dispatched, not after
No exceptions for simple or single-line fixes. A one-line change can introduce a SQL injection, missing escape, or auth bypass just as easily as a large feature. Simplicity of the change is not a reason to skip this skill.
Subagent Dispatch Rules
When using subagent-driven development (superpowers:subagent-driven-development or equivalent) on a WordPress plugin or theme:
Invoke wp-compliance once, in the controller, before dispatching Task 1. Do not defer it to the implementer.
Include a WP Compliance section in every implementer subagent prompt, pasting the Pre-Code Checklist and the Safe Default Order verbatim:
## WP Compliance (mandatory)
Run through this checklist before writing any code:
- Every PHP file starts with an ABSPATH guard (`defined( 'ABSPATH' ) || exit;`)
- All $_GET/$_POST/$_REQUEST/$_COOKIE/$_SERVER/REST/AJAX input validated at entry
- Output escaped at render time, context-specific (esc_html, esc_attr, esc_url, wp_kses)
- Capability check in place for every privileged action
- Nonce verified for state-changing requests — paired with capability check, never alone
- No raw variables in SQL — $wpdb->prepare() for values, allowlists for column/direction names; LIKE wildcards passed as %s via $wpdb->esc_like()
- Sanitizer is flat and outermost (e.g. `absint( wp_unslash( $_GET['x'] ?? 0 ) )`), not nested inside trim()/strtolower()/etc.
- JSON input: wp_unslash → json_decode → sanitize decoded values (do NOT sanitize_text_field before decode)
- Structured POST/REST maps (nested arrays from bracket-syntax forms or FormData): walk + sanitize per-value, not just outer `(array) wp_unslash`
- No hardcoded API keys, tokens, or credentials
- No eval(), unserialize() on untrusted data, or dynamic includes from user input
- Release files have consistent line endings (no Plugin Check `Internal.LineEndings.Mixed`)
Safe order: validate input → sanitize when needed → check capability → verify nonce → perform action safely → escape output late
Reviewers (spec compliance and code quality) must also receive this checklist in their prompt so they can flag violations the implementer missed.
Failing to include compliance context in a subagent prompt is the same P10 violation as not invoking the skill at all — the subagent writes code without safety constraints.
Pre-Code Checklist
Run through this before writing any code. Every item must be addressed:
Every PHP file starts with defined( 'ABSPATH' ) || exit; (or equivalent guard)
All $_GET, $_POST, $_REQUEST, $_COOKIE, $_SERVER, REST/AJAX input validated at entry
Output escaped at render time, context-specific (esc_html, esc_attr, esc_url, wp_kses, etc.)
Capability check in place for every privileged action
Nonce verified for state-changing requests — paired with a capability check, never alone
No raw variables in SQL — $wpdb->prepare() used for values, allowlists for column/direction names, LIKE wildcards passed via $wpdb->esc_like() + %s
Sanitizer is flat and outermost (absint( wp_unslash( ... ) )), not nested through trim()/strtolower()/etc.
Structured POST/REST maps (nested arrays from bracket-syntax forms or FormData) walked + sanitized per-value — outer (array) wp_unslash is not sufficient
No hardcoded API keys, tokens, or credentials
No eval(), unserialize() on untrusted data, or dynamic includes from user input
Release files have consistent line endings (no Plugin Check Internal.LineEndings.Mixed)
Rules
1 — Input & Output
1. Do not trust any input.
Do not trust $_GET, $_POST, $_REQUEST, $_COOKIE, REST input, AJAX payloads, shortcode attrs, uploaded files, third-party API responses, or values already stored in your own database. Validate first, sanitize when needed. WordPress guidance explicitly favors validation over sanitization when you can define allowed values.
2. Do not sanitize once and assume it is safe forever.
A value can be valid for storage but unsafe for HTML, attributes, JavaScript, URLs, SQL, or JSON output. Sanitizing input and escaping output are separate steps with different purposes. Do not reuse a "cleaned" value across contexts.
3. Do not skip escaping on output.
Escape as late as possible and for the exact output context. This applies to admin pages, settings screens, notices, metaboxes, REST responses rendered in JS templates, and frontend markup. WordPress explicitly recommends escaping right before rendering.
15. Do not output raw HTML unless you truly control it.
If HTML is allowed in a settings field, apply a strict allowed-tags policy via wp_kses. For plain text settings, keep them plain text. Do not save and reprint arbitrary HTML from user-controlled fields.
2 — Auth & Access Control
4. Do not rely on nonces as permission checks.
A nonce is not authorization. check_admin_referer() and check_ajax_referer() protect against CSRF — they do not decide whether a user is allowed to act. Always pair nonce checks with capability checks.
5. Do not skip capability checks.
Do not let any logged-in user hit privileged actions just because they can access wp-admin, an AJAX endpoint, or a REST route. Check the exact capability required for the action. This applies to settings saves, imports, exports, deletes, remote requests, and debug tools.
11. Do not make privileged AJAX or REST routes too open.
Every state-changing route must verify a nonce where relevant and check capabilities explicitly. Do not assume "admin page only" means the endpoint is safe — attackers call endpoints directly.
12. Do not use broad, weak capabilities.
Do not default to manage_options for everything. Use the smallest capability that matches the action. Do not grant dangerous actions to lower roles for convenience.
16. Do not assume admin-only means safe.
Admin pages still process untrusted input, can suffer CSRF without nonces, and can expose XSS if output is not escaped. Apply the same security rules on both admin and public sides.
3 — Database
6. Do not build SQL with raw variables.
Do not concatenate untrusted values into SQL — including WHERE, ORDER BY, LIMIT, IN (...), table names, meta keys, or search fragments. Use $wpdb->prepare() for values. Whitelist allowed columns and sort directions for sortable fields.
7. Do not accept "sanitized enough" SQL fragments.
Do not pass around a prebuilt $where_clause, $orderby, or $limit string assembled from request values. Even if parts look cleaned, they are easy to get wrong. Build query parts from fixed templates and strict allowlists. This includes fragments individually pre-escaped with $wpdb->prepare() — the correct pattern is to keep SQL templates and values in separate arrays, then call $wpdb->prepare() once on the complete query with all values merged.
19. Do not build dynamic WHERE clauses for Plugin Check-compatible code.
Plugin Check's static analyzer cannot follow dynamically assembled $where_sql variables, even when properly prepared — it will flag {$where_sql} as InterpolatedNotPrepared and UnescapedDBParameter regardless. The correct pattern is static SQL templates with opt-out bypass conditions:
// Instead of: $where_sql = implode(' AND ', $conditions); ... WHERE {$where_sql}
// Use a static template with bypass per optional filter:
$wpdb->prepare(
"WHERE (%s = '' OR r.match_type = %s) AND (%s = '' OR r.asset_type = %s) LIMIT %d",
$match_type, $match_type, $asset_type, $asset_type, $limit
);
// Pass '' for inactive filters ('' = '' is TRUE, condition bypassed).
// Pass the real value twice for active filters (first arg fails, second filters).
This gives a fixed argument count, no interpolated variables, and full static verifiability. (user flagged 2026-04-10)
4 — Files, Uploads & Remote Requests
8. Do not allow unsafe file uploads.
Do not trust file names, browser MIME types, file extensions, or image metadata. Do not let users upload PHP-capable files, SVGs, archives, or arbitrary documents without a safe handling path. Validate tightly. Storage paths must not allow execution.
13. Do not allow arbitrary remote requests.
Do not let request parameters decide remote URLs, webhook targets, download locations, or proxy destinations — this enables SSRF and data exfiltration. If the plugin talks to external services, use a strict domain allowlist and safe request patterns.
14. Do not eval, unserialize, or execute user-controlled content.
Do not use eval(), dynamic includes from user input, shell execution functions, or unserialize() on untrusted data. These are high-risk patterns that can get a plugin rejected or pulled if exploited.
5 — Secrets, Logs & Cleanup
9. Do not expose debug data to normal users.
Do not leave debug endpoints, diagnostic pages, raw logs, headers, stack traces, SQL errors, tokens, or system paths visible to subscribers, customers, or anonymous visitors. Protect with capabilities and keep out of public web paths.
10. Do not store secrets carelessly.
Do not hardcode API keys, license secrets, or private tokens. Do not print them into HTML, localize them to JS unless absolutely necessary, or expose them in error messages. Keep sensitive values minimal, access-controlled, and out of logs.
17. Do not ignore uninstall and cleanup security.
Do not leave behind unsafe options, cron jobs, temp files, logs, or custom tables with sensitive data after uninstall. Uninstall routines must include proper permission checks.
18. Do not hide problems with phpcs ignores if the code is actually unsafe.
Do not silence Plugin Check or PHPCS warnings unless the warning is a genuine false positive you can justify. The real fix is usually the right fix.
20. Document phpcs:ignore only for genuine false positives — and justify each one.
PHPCS cannot trace through helper methods, exception propagation paths, or include-scope. Common genuine false positives:
ExceptionNotEscaped when the exception is caught and passed to wp_send_json_error() (not rendered as HTML)
NonceVerification.Missing when nonce is verified in a shared helper (e.g. $this->check())
NonPrefixedVariableFound for variables in template files included within class methods (local scope, not global)
NonceVerification.Recommended for token-based frontend endpoints where nonces are architecturally incompatible (e.g. headless browser requests)
WordPress.PHP.DevelopmentFunctions.error_log_error_log when error_log() is used for intentional production server-side logging of exception details that are explicitly withheld from the browser response (i.e. the paired wp_send_json_error() call uses a generic message, not $e->getMessage())
WordPress.DB.DirectDatabaseQuery.SchemaChange (and the accompanying DirectQuery + NoCaching) for DDL statements (ALTER TABLE … MODIFY COLUMN, CREATE TABLE, DROP TABLE) inside a version-gated maybe_upgrade() migration block. dbDelta() cannot handle MODIFY COLUMN on ENUM or other structural changes; a direct $wpdb->query() is the only correct tool. Always gate with version_compare() to keep the block idempotent. (flagged 2026-04-17)
WordPress.WP.AlternativeFunctions.file_system_operations_{fopen,fclose,fwrite} when the target is a PHP stream wrapper (php://memory, php://output, php://temp, php://input). WP_Filesystem operates on real filesystem paths and has no equivalent for stream wrappers — patterns like buffering CSV into php://memory before ZipArchive::addFromString(), or streaming response bodies via php://output, cannot go through it. Justify each instance with the specific stream and purpose. (flagged 2026-04-24 after Plugin Check audit)
WordPress.WP.AlternativeFunctions.file_system_operations_readfile when streaming a server-generated temp file (from wp_tempnam() or similar) directly to the HTTP response body as a binary download. is the pragmatic choice for binary pass-through — loading via would blow memory on large archives, and WP has no helper for this pattern. Justify with the source of the path and why memory-loading isn't viable.
Always add a trailing comment explaining it is safe:
PHPCS suppression playbook — placement mechanics
Before suppressing anything, try to fix it. Most sniffs warn about patterns that should change — use $wpdb->prepare() with placeholders, escape output with the context-specific esc_* function, flatten nested sanitizers (Rule 24), refactor dynamic WHERE to static template with bypass conditions (Rule 19), collapse a multi-line statement that doesn't need to be multi-line. Suppression is a last resort, only for genuine false positives with a clear human-readable justification.
When suppression is the right call, placement matters. These are the lessons from real audits where wrong placement forced multiple iterations before Plugin Check came back clean. (flagged 2026-04-23 after Plugin Check iteration cycle)
1. Match the directive to the statement shape.
Statement shape
Directive
Notes
Single-line statement with all flagged sniffs on that one line
// phpcs:ignore <sniffs> -- reason directly above
Simplest; cheapest; prefer this.
Multi-line statement where sniffs fire on different lines (multi-line SQL string with {$table}, prepare() with spread args, ternary/conditional across lines)
// phpcs:disable <sniffs> -- reason before + // phpcs:enable <sniffs> after
Covers every line between the two directives.
Multi-line statement that can be collapsed without hurting readability
Collapse to single line + one phpcs:ignore above
Often cleanest — no closing directive, scope auto-correct.
2. Critical scope rule: phpcs:ignore applies to the first line of the next statement only.
Sniffs fire per-line, not per-statement. Different sniffs within the same statement can land on different lines. A single-line phpcs:ignore above a multi-line statement covers the FIRST line of the statement — not lines 2, 3, 4 of the same SQL string.
// Wrong — phpcs:ignore applies to line with $wpdb->query( ; sniff fires on string line, not covered.
// phpcs:ignore WordPress.DB.PreparedSQL.InterpolatedNotPrepared -- table name from plugin constant.
$wpdb->query(
"ALTER TABLE `{$t}` ADD COLUMN `foo` INT"
);
// Fix A — collapse so one annotation covers the whole call:
// phpcs:ignore WordPress.DB.PreparedSQL.InterpolatedNotPrepared -- table name from plugin constant.
$wpdb->query( "ALTER TABLE `{$t}` ADD COLUMN `foo` INT" );
// Fix B — disable/enable brackets cover every line in between:
// phpcs:disable WordPress.DB.PreparedSQL.InterpolatedNotPrepared -- table name from plugin constant.
$wpdb->query(
"ALTER TABLE `{$t}` ADD COLUMN `foo` INT"
);
// phpcs:enable WordPress.DB.PreparedSQL.InterpolatedNotPrepared
Stacked phpcs:ignore directives don't chain — the second consumes the first's scope before the target statement is reached. Two consecutive phpcs:ignore lines above a single statement apply only the SECOND one — the first's one-line scope is spent on the second phpcs:ignore line. Combine all sniffs into a comma-separated list on one annotation, or use phpcs:disable/phpcs:enable brackets if you want independent justifications per sniff cluster.
// Wrong — first directive's scope is spent on the second directive line; DirectQuery/NoCaching NOT suppressed on the target.
// phpcs:ignore WordPress.DB.DirectDatabaseQuery.DirectQuery,WordPress.DB.DirectDatabaseQuery.NoCaching -- custom plugin table.
// phpcs:ignore WordPress.DB.PreparedSQL.InterpolatedNotPrepared -- table name from constant.
$count = (int) $wpdb->get_var( "SELECT COUNT(*) FROM `{$t}`" );
// Right — all three sniffs on one annotation line, single combined justification.
// phpcs:ignore WordPress.DB.DirectDatabaseQuery.DirectQuery,WordPress.DB.DirectDatabaseQuery.NoCaching,WordPress.DB.PreparedSQL.InterpolatedNotPrepared -- custom plugin table; table name from constant, not user input.
$count = (int) $wpdb->get_var( "SELECT COUNT(*) FROM `{$t}`" );
Two stacked annotations can look fine to php -l and to a human reviewer — the bug surfaces only in the next Plugin Check run, which still fires the first annotation's sniffs despite the visible directive. (flagged 2026-04-24 after Plugin Check audit.)
3. Name every sniff that can fire inside the suppressed scope. Partial sniff lists leak warnings. Common clusters observed:
Pattern
Sniffs to list
Custom-table $wpdb->get_var/update/insert/delete (no prepare())
Bare $wpdb->query($sql) where $sql is a variable holding a literal
add WordPress.DB.PreparedSQL.NotPrepared
$_GET/$_POST/$_SERVER reads on admin-list filters
WordPress.Security.NonceVerification.Recommended (GET) OR .Missing (POST) — pick the exact variant
4. No blank lines between the directive and the statement. A blank line consumes the directive's scope — it applies to the blank line and stops there. Keep the directive immediately above the target.
5. If you're listing more than 3 different sniffs on one statement, that's a signal. Consider whether the code itself should be refactored instead. Three or more sniff IDs on one annotation often means the code is doing too much in one place; breaking it up (or rewriting per Rule 19's static-template pattern) usually produces cleaner code and eliminates the need to suppress.
6. Verify file-by-file after each batch. Running Plugin Check once after touching 20 files and seeing the same warnings reappear wastes time. Iterate in groups of 3–4 files, re-run the sniff locally or on CI, confirm before moving on.
7. Every suppression needs a per-instance justification.-- false positive alone is not enough. Explain why the code is safe and how the static sniff is mis-reading it. Future reviewers need the justification to evaluate whether it still holds after refactors.
8. phpcs:enable is static — control-flow branches do not scope it. An enable placed inside if/early-return closes the surrounding disable for every line below in the file, not just the branch. If you need to suppress through an early return, either (a) delete the in-branch enable and let the outer enable close the scope, or (b) split the code into two disable/enable brackets — one covering the pre-return statement, one covering the post-return statements. (flagged 2026-04-23 after Plugin Check audit)
21. Every plugin PHP file must abort when loaded directly.
Add defined( 'ABSPATH' ) || exit; (or if ( ! defined( 'ABSPATH' ) ) exit;) at the very top of every PHP file — main bootstrap, class files, REST handlers, AJAX handlers, template partials, trait files, autoloaded files, include-only helpers. No exceptions. Plugin Check reports missing_direct_file_access_protection as an ERROR. A misconfigured web server that serves raw .php files from the plugin directory (Apache without mod_php wired correctly, a misrouted nginx location block, a file copied to a debug path) will dump file contents to anyone — leaking class structure, SQL templates, and occasionally secrets from development. Cost: one line. Value: zero information disclosure when the server misbehaves. (flagged 2026-04-22 after Plugin Check audit)
22. Parameterize LIKE wildcards; never hardcode them inside a prepared query.
Pass the wildcard pattern as a %s parameter built from $wpdb->esc_like(). Never write LIKE 'prefix.%' inside a $wpdb->prepare() template.
// Wrong — Plugin Check flags LikeWildcardsInQuery; and if $prefix is ever user input,
// unescaped % or _ become wildcard injection.
$wpdb->prepare(
"SELECT * FROM `{$t}` WHERE slug LIKE 'prefix.%%' AND created_at < %s",
$cutoff
);
// Right — wildcard is a parameter, literal portion is esc_like'd.
$pattern = $wpdb->esc_like( 'prefix.' ) . '%';
$wpdb->prepare(
"SELECT * FROM `{$t}` WHERE slug LIKE %s AND created_at < %s",
$pattern,
$cutoff
);
Two reasons. First: esc_like() escapes literal % and _ within the prefix so user-supplied prefixes can't inject wildcards. Second: the static sniff can verify a fully-placeholdered query but not a hardcoded wildcard, so hardcoding fails Plugin Check. (flagged 2026-04-22 after Plugin Check audit)
23. $_SERVER is untrusted input — treat HTTP_ / REDIRECT_ / REMOTE_* values like $_GET.**
Rule 1 lists the obvious sources. $_SERVER belongs there too. Any key whose value originates in the HTTP request (HTTP_* headers, REDIRECT_* from mod_rewrite forwards, REMOTE_ADDR behind an untrusted proxy, QUERY_STRING, REQUEST_URI, HTTP_REFERER, HTTP_USER_AGENT) is attacker-controllable. Sanitize and wp_unslash before use.
// Common Apache/FastCGI fallback for Authorization header passthrough.
// Wrong — raw value assigned to HTTP_AUTHORIZATION for downstream consumers.
if ( ! isset( $_SERVER['HTTP_AUTHORIZATION'] ) && isset( $_SERVER['REDIRECT_HTTP_AUTHORIZATION'] ) ) {
$_SERVER['HTTP_AUTHORIZATION'] = $_SERVER['REDIRECT_HTTP_AUTHORIZATION'];
}
// Right — sanitize + unslash, same as $_POST.
if ( ! isset( $_SERVER['HTTP_AUTHORIZATION'] ) && isset( $_SERVER['REDIRECT_HTTP_AUTHORIZATION'] ) ) {
$_SERVER['HTTP_AUTHORIZATION'] = sanitize_text_field( wp_unslash( $_SERVER['REDIRECT_HTTP_AUTHORIZATION'] ) );
}
(flagged 2026-04-22 after Plugin Check audit)
24. Sanitizer placement must be recognizable to the static sniff.
Plugin Check's ValidatedSanitizedInput.InputNotSanitized sniff does not trace through nested calls beyond one level and does not recognize type casts as sanitizers. Code that is semantically safe but violates these patterns will still be flagged as ERROR. Prefer flat, canonical patterns with a recognized WordPress sanitizer as the outermost wrapper:
// Wrong — (int) cast is not a recognized sanitizer prefix; nested wp_unslash inside trim() inside sanitize_url() defeats the sniff.
$page = (int) $_GET['paged'];
$url = sanitize_url( trim( wp_unslash( $_POST['callback_url'] ?? '' ) ) );
// Right — outermost call is a recognized sanitizer; wp_unslash is the inner step; no intermediate helpers.
$page = absint( wp_unslash( $_GET['paged'] ?? 0 ) );
$raw = wp_unslash( $_POST['callback_url'] ?? '' );
$url = sanitize_url( trim( $raw ) );
Rule of thumb: read + unslash + sanitize on one or two adjacent lines; the outermost function call must be a recognized WordPress sanitizer (sanitize_text_field, sanitize_email, sanitize_url, esc_url_raw, absint, intval, sanitize_key, wp_kses, wp_kses_post, sanitize_textarea_field, etc.). Do not rely on (int) / (string) / (bool) casts alone — they're semantically safe but invisible to the sniff. (flagged 2026-04-22 after Plugin Check audit)
25. JSON input: decode BEFORE sanitizing; sanitize each decoded value, not the raw JSON string.sanitize_text_field strips line breaks, collapses whitespace, removes certain escape characters, and mangles embedded quotes — all of which corrupt JSON and make json_decode return null or lose data. The correct order for JSON request bodies is:
wp_unslash to reverse WordPress's magic-quote emulation.
json_decode( ..., true ) to get an associative array.
Type-check the result (must be is_array).
Sanitize each decoded value according to its expected shape.
This also applies to JSON bodies in REST handlers ($request->get_json_params() returns already-decoded values — but if you read raw body via $request->get_body(), same rule). (flagged 2026-04-22 after Plugin Check audit)
26. Prefer wp_delete_file() over bare or @-suppressed unlink() for file cleanup.
Plugin Check flags raw unlink() and @unlink() as WordPress.WP.AlternativeFunctions.unlink_unlink. wp_delete_file() is a thin core wrapper that runs @unlink() through the wp_delete_file filter, so it is a drop-in replacement that passes the sniff. Use it for any filesystem cleanup — temp files from wp_tempnam(), stale caches, uninstall residue.
// Wrong — flagged by Plugin Check; forces a phpcs:ignore.
@unlink( $tmp_path );
// Right — WP-idiomatic, filter-aware, no suppression needed.
wp_delete_file( $tmp_path );
Caveat:wp_delete_file() returns void. If you need to detect cleanup failure to log it (e.g. diagnosing temp-dir drift or permissions issues), you must keep @unlink() with a phpcs:ignore WordPress.WP.AlternativeFunctions.unlink_unlink -- need return value for debug-only error_log on cleanup failure. annotation. In practice this is rarely worth the extra suppression — silent cleanup plus occasional orphaned temp files is usually acceptable. (flagged 2026-04-24 after Plugin Check audit)
27. Structured POST/REST maps: sanitize every value, not just the outer container.
When $_POST['x'] carries a multi-level array (typical form syntax x[key][]=val or fetch-API FormData with bracket keys), (array) wp_unslash( $_POST['x'] ) only unslashes the outer level. The inner-level scalars are still untrusted strings.
Walk the structure with explicit loops, validate each key, validate the shape of each value, and apply a recognized WordPress sanitizer or an allowlist regex to every leaf scalar before using it. Apply the same shape check that JSON bodies get (Rule 25) — the only difference is PHP did the decode for you instead of json_decode.
PHP's $_POST parser turns option_map[key1][]=foo&option_map[key1][]=bar into a real nested array. The outer (array) wp_unslash recursively unslashes magic-quoted strings — but it does not enforce structure, validate keys, or sanitize values. That is your job, the same way it is for json_decode'd JSON bodies (Rule 25). The two patterns differ only in who does the decode: PHP for bracket-syntax forms, your code for raw JSON. The trust-no-input + per-value-sanitize discipline is identical.
If the inner values must conform to a specific shape (a known enum, a known character class, a known max length), validate-not-sanitize: an allowlist regex or in_array( $v, $allowed, true ) is stronger than sanitize_text_field, which only strips a few specific patterns and lets through values that pass-the-strip but still don't belong. This matters more for structured maps than for flat scalars, because the per-leaf trust budget is multiplied across every key/value the operator can submit. (flagged 2026-05-15 after wp-compliance audit during a structured-POST-map-sanitization audit)
28. Normalize line endings before Plugin Check/release.
Plugin Check reports Internal.LineEndings.Mixed when a file contains both CRLF and LF line endings. Treat this as a release-blocking hygiene issue, not a false positive. After touching plugin PHP/readme/assets, run a line-ending check on changed files and normalize each affected file to one style before claiming Plugin Check is clean. This especially matters after edits from mixed tools (PowerShell, patch tools, IDEs) because php -l and JS syntax checks can still pass while Plugin Check warns. (flagged 2026-06-29 after Plugin Check report)
Safe Default Order
Every WordPress action should follow this sequence:
validate input → sanitize when needed → check capability → verify nonce → perform action safely → escape output late
Quick Release Checklist
Before releasing or committing, confirm you are NOT:
Shipping any PHP file without an ABSPATH guard at the top
Echoing raw data anywhere
Using request values in SQL
Hardcoding LIKE wildcards inside a prepared query
Saving settings without capability + nonce checks
Sanitizing raw JSON before json_decode (corrupts the payload)
Outer-cast-only sanitization on structured POST/REST maps (inner scalars left untrusted)
Reading $_SERVER HTTP_/REDIRECT_/REMOTE_* values without sanitize + wp_unslash
Exposing debug pages or logs
Trusting uploads, remote URLs, or API responses
Relying on sanitization alone instead of validation + escaping
Ignoring Plugin Check warnings without justification
Shipping files with mixed CRLF/LF line endings (Internal.LineEndings.Mixed)
When Reviewing a Plugin Check / PHPCS Report
When the user provides a list of Plugin Check, PHPCS, or other static-analyzer errors and asks you to address them:
Triage every error against the numbered rules to classify it as:
Real bug — fix per the relevant rule.
False positive — annotate with phpcs:ignore Sniff.Name -- reason per Rule 20. Include the justification inline.
Out of scope — doc/license/branding or non-security issues the user explicitly excluded; skip.
Fix the real bugs. Prefer the smallest surgical edit. Run the test suite after each logical batch to catch regressions.
Run a meta-check against the skill itself. After triage, ask:
Is any existing rule written too loosely? A concern that should have been caught by Rule N but the wording left ambiguity → tighten Rule N.
Is there a recurring error category not cleanly covered by any rule? → new rule worth adding.
Is there a false-positive pattern seen more than once? → add an example to Rule 20's list.
End the report with a short status line as a bullet list summarizing: how many valid errors were fixed, how many were annotated as false positives, how many skill-level changes surfaced. Ask the user whether to append/tighten the skill. Example:
17 valid errors fixed across 6 commits
41 false positives annotated with phpcs:ignore + justification
2 new rules worth adding:
ABSPATH guard on every PHP file
$_SERVER as untrusted input
1 existing rule to tighten:
Rule 6 should name LIKE wildcards explicitly
Append to skill now? (local-only by default; remote push only on explicit YES)
NEVER auto-apply skill changes. No matter how obvious the new rule seems, always ask for explicit user confirmation before editing SKILL.md. The user may want to tweak wording, scope it differently, or decline entirely. An unsolicited skill edit is a policy change without review.
Local-commit default; remote push requires explicit YES. If the user says "yes, add it" or similar, that authorizes only the local file edit + local commit. Pushing to any remote (GitHub mirror, compliance backup, team fork) requires a separate explicit YES — this skill lives on a public repo and every push is public record.
This closes the feedback loop so the skill improves with each real-world audit instead of staying frozen. The meta-check is mandatory — do not hand back a "done" report without it.
Appending New Rules
When the user identifies a new security concern or coding issue:
Add it as a numbered rule (continue the numbering: 19, 20, …)
Place it under the most appropriate category above
Update the pre-code checklist if the new rule warrants a mandatory check
Update the quick release checklist if it applies at review time
Note the source (e.g., "user flagged 2026-04-10")
readfile()
file_get_contents()
(flagged 2026-04-24 after Plugin Check audit)
why
// phpcs:ignore Sniff.Name -- reason.
(flagged in audit 2026-04-11)
Sanitize every example you add to the skill before writing it to disk. The skill file is publicly shared. When adding rules, examples, checklist items, or appendix entries:
No project names — use "my plugin", "the plugin", "your plugin".
No real table names — use my_plugin_table, {$t}, {$prefix}_my_table.
No real file paths — use includes/class-example.php, admin/class-settings.php.
No commit SHAs, PR numbers, branch names, or tag names.
No customer identifiers, user IDs, API keys, token shapes, endpoint URLs, domain names, or internal service secrets.
No direct quotes from proprietary code — rewrite as a generic pattern that illustrates the same concept.
Date-stamp the source only generically — *(flagged 2026-04-22 after Plugin Check audit)* is fine; *(flagged during the XYZ Corp audit)* is not.
Before saving the edit, re-read every example and checklist line and strip anything that identifies a specific project, repo, customer, or private system.
