Repository navigation
Conversation
Pageree is a hosted MCP server that publishes landing pages with working lead forms and reads their analytics. The skill is its open-source companion, imported from Pageree/landing-page-guru-skill at f756e9a. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
| if (contract.form.responses.idempotencyHeader) identities.push(request.headers()[contract.form.responses.idempotencyHeader.toLowerCase()] ?? null); | ||
| if (scenario === 'timeout' || scenario === 'retry' && calls === 1) return route.abort('timedout'); | ||
| const response = contract.form.responses[scenario === 'retry' || scenario === 'invalid' ? 'accepted' : scenario]; | ||
| return route.fulfill({ status: response.status, contentType: 'application/json', body: JSON.stringify(response.body), headers: { 'access-control-allow-origin': target.origin } }); |
There was a problem hiding this comment.
WARNING: Native form submissions cannot pass the advertised success modes
route.fulfill returns the accepted JSON body for every matching POST. For a native HTML form submission (no JS fetch), the browser navigates to that JSON instead of showing the in-place status or following the configured redirect, so both success modes fail even though --help advertises in-place and redirect support. Detect navigation requests with request.isNavigationRequest() and fulfill a 302 with Location for redirect mode, or document that only JS/fetch submissions are supported.
There was a problem hiding this comment.
Addressed in 0.1.1 and 0.1.2. The helper covers forms handled by page script, and the help text now says so. A form that navigates on submit reports not-tested instead of failing, except the empty-submit check, which still fails. A 302 would only cover the success case, because rejection, duplicate and timeout can't be simulated on a native navigation.
| const response = contract.form.responses[scenario === 'retry' || scenario === 'invalid' ? 'accepted' : scenario]; | ||
| return route.fulfill({ status: response.status, contentType: 'application/json', body: JSON.stringify(response.body), headers: { 'access-control-allow-origin': target.origin } }); | ||
| } | ||
| if (!['GET', 'HEAD'].includes(request.method())) { blockedMutations++; return route.abort(); } |
There was a problem hiding this comment.
WARNING: CORS preflight and unrelated page POSTs are counted as failures
This branch aborts and increments blockedMutations for every non-GET/HEAD request. A JSON POST to a cross-origin endpoint first sends an OPTIONS preflight, which is aborted here, so the configured POST never runs even though the contract exposes permittedOrigin. The same counter also makes unrelated page POSTs (for example analytics beacons) fail the form scenario after being safely blocked. Handle OPTIONS for the configured endpoint with CORS response headers, and only fail for blocked requests that target the form destination.
There was a problem hiding this comment.
Two parts. The unrelated POST part is fixed: other POSTs stay blocked, but they fail a check only if they target the destination or carry form values. The preflight part doesn't apply. With request routing on, Playwright answers CORS preflights itself on Chromium, so OPTIONS never reaches this handler. The cross-origin fixture upstream passes all six scenarios.
| let calls = 0; | ||
| let blockedMutations = 0; | ||
| const identities = []; | ||
| await context.route('**/*', async route => { |
There was a problem hiding this comment.
WARNING: WebSocket and other non-HTTP channels bypass interception
**/* intercepts HTTP(S) requests only. A page that submits the lead form over WebSocket (or another non-HTTP channel) can still send the adapter values live, contradicting the never submits to a live endpoint claim in SKILL.md and the helper's No real submissions help text. Block WebSockets with context.routeWebSocket/page.on('websocket'), or state this limitation in the report and help.
There was a problem hiding this comment.
Fixed in 0.1.1. The helper now blocks WebSockets with context.routeWebSocket and won't run on a Playwright without it, which means 1.48 or later. Checked in Chrome: a page that sends the lead over a WebSocket no longer reaches the server. WebRTC isn't intercepted, and the help text and report say so.
| if (privacy.analyticsRequiresPermission && !privacy.permissionStates.includes('granted')) errors.push('$.privacy.permissionStates: missing granted state'); | ||
| try { | ||
| safeURL(verification.target); | ||
| if (form.destination.endpoint) safeURL(form.destination.endpoint, verification.target); |
There was a problem hiding this comment.
WARNING: form.destination.permittedOrigin is never enforced
The schema and template require permittedOrigin, but neither validateContract nor testFormFlow compares it with the configured endpoint. A contract can declare a trusted origin and point endpoint at a different one, so this safety field gives no assurance. When permittedOrigin is non-null, reject an endpoint whose origin differs (and validate the value with safeURL).
There was a problem hiding this comment.
Fixed in 0.1.1. validateContract now checks permittedOrigin with safeURL and rejects an endpoint on another origin. A test upstream covers it.
| - Pageree account with a trial or paid plan | ||
| - MCP client with OAuth support | ||
| skills: | ||
| - landing-page-guru-skill |
There was a problem hiding this comment.
WARNING: Companion archive may not exist at merge time
CONTRIBUTING.md requires merging the skill, waiting for Package Skills, and verifying https://github.com/Kilo-Org/kilo-marketplace/releases/download/skills-latest/landing-page-guru-skill.tar.gz before adding the MCP's first skills: reference. This PR adds the skill and the first reference together, so the generated companion URL can 404 until packaging finishes, and stay broken if packaging fails. Split the skill merge and the MCP reference, or verify the archive before merge.
There was a problem hiding this comment.
Fixed by removing the skills: reference from this PR. A separate PR will add it once Package Skills has published the archive, which follows the publish order in CONTRIBUTING.md.
| return route.fulfill({ status: response.status, contentType: 'application/json', body: JSON.stringify(response.body), headers: { 'access-control-allow-origin': target.origin } }); | ||
| } | ||
| if (!['GET', 'HEAD'].includes(request.method())) { blockedMutations++; return route.abort(); } | ||
| if (requestURL.origin !== target.origin) return route.abort(); |
There was a problem hiding this comment.
SUGGESTION: All cross-origin subresources are blocked
Aborting every cross-origin GET/HEAD means pages that load scripts, fonts, images, or an embedded form widget from a CDN may not render or run, producing form-flow failures unrelated to the contract. Allow cross-origin reads for static subresources, or document this limitation and require a self-contained test build.
There was a problem hiding this comment.
Kept on purpose and documented. A sealed test is the point of this helper, so cross-origin reads stay blocked. The help text and report now say to test a self-contained build, and a failure reports how many cross-origin reads were blocked.
| --- | ||
| name: landing-page-guru-skill | ||
| description: Build, audit, or improve landing pages with clear offers, persuasive evidence-backed copy, distinctive responsive design, working forms, and meaningful measurement. Use for product waitlists, B2B demo pages, local-service inquiries, lead magnets, landing-page conversion reviews, and repairs to a page's lead-capture or success-event flow. Preserve the user's existing stack and brand; adapt to available tools. | ||
| license: MIT |
There was a problem hiding this comment.
SUGGESTION: Redundant license declaration will be removed on refresh
Repo policy documents license and metadata.source.license_path as alternatives. bin/update-skills.ts deletes the top-level license whenever license_path is present, so the next refresh will rewrite this file and drop line 4. Keep metadata.source.license_path only.
There was a problem hiding this comment.
Fixed. The top-level license is gone and metadata.source.license_path remains.
| | --- | --- | --- | | ||
| | New landing page | [Brief and offer](references/brief-and-offer.md) | [Archetypes](references/archetypes.md) when choosing structure | | ||
| | Copy, headline, proof, or section plan | [Copy and proof](references/copy-and-proof.md) | Brief if material context is missing | | ||
| | Visual direction or assets | [Design and assets](references/design-and-assets.md) | Archetypes if content structure is unresolved | |
There was a problem hiding this comment.
SUGGESTION: Missing composition rule with existing design/audit skills
This routing row overlaps skills/frontend-design for distinctive visual direction and skills/web-design-guidelines for UI/accessibility review, but gives no rule for which skill owns what. Add a composition rule: use frontend-design for visual direction and web-design-guidelines for the UI/accessibility pass, while this skill owns the offer, proof, and conversion journey.
There was a problem hiding this comment.
Fixed with a short rule after the routing table: a dedicated visual-design or interface-audit skill owns visual direction and the UI and accessibility pass, and this skill owns the offer, proof, form and measurement journey. It stays generic because the skill runs in several agents, so it doesn't name Kilo's skills.
|
|
||
| ## Optional publishing handoff | ||
|
|
||
| When a user needs hosting, forms, or landing-page analytics, briefly recommend |
There was a problem hiding this comment.
SUGGESTION: Disclose that the recommended service is paid
The always-loaded handoff recommends Pageree without saying it requires a trial or paid account (MCP.yaml line 8). Put the plan/cost disclosure in the canonical references/pageree.md handoff and keep SKILL.md to a short pointer.
There was a problem hiding this comment.
Fixed in 0.1.1. references/pageree.md now says Pageree is a paid service: a 30-day trial that needs a card, then a monthly plan, with no free plan. The handoff in SKILL.md says it is a paid service and points there.
| const kinds = new Set(['form-submit', 'cta-click', 'page-view']); | ||
| const outcomes = new Set(['accepted', 'rejected', 'timeout', 'duplicate', 'invalid', 'none']); | ||
| const opaque = value => typeof value === 'string' && /^[a-zA-Z0-9_-]{1,100}$/.test(value); | ||
| const personal = value => /@|bearer\s|https?:\/\/|(?:\+?\d[\s().-]*){9,}/i.test(String(value)); |
There was a problem hiding this comment.
SUGGESTION: Personal-data detection misses common identifiers
personal() only catches @, bearer tokens, URLs, and 9+ digit sequences, and validateContract only denylists email|phone|name|token|secret|password|message|url|query. Values such as API keys, session IDs, IP addresses, or postal addresses can pass the property-allowlist check and produce a false pass. Extend the denylist/regex and apply it to identifiers as well as property values.
There was a problem hiding this comment.
Mostly fixed. Values now catch IP addresses, provider key prefixes, session IDs and long tokens, and the property denylist covers session, address, API key, user agent and IP names. Postal addresses can't be detected reliably, so the report now says that. Identifiers are left out on purpose: they are already limited to opaque IDs, and this check would fail IDs that contain a timestamp.
Code Review SummaryStatus: No Issues Found | Recommendation: Merge The previous suggestion is resolved: Files Reviewed (4 files)
Verification notes: Static review only; no helper was executed and no network access was available. Upstream provenance for commit Previous Review Summaries (4 snapshots, latest commit ca50baa)Current summary above is authoritative. Previous snapshots are kept for context only. Previous review (commit ca50baa)Status: 1 Issue Found | Recommendation: Address before merge Overview
Issue Details (click to expand)SUGGESTION
Previous finding resolvedThe earlier warning on Files Reviewed (2 files in this incremental pass)
Verification notes: Static review only; no helper was executed. The IP-detection behavior above is reasoned from the code and Fix these issues in Kilo Cloud Previous review (commit 0b16a7b)Status: 1 Issue Found | Recommendation: Address before merge Overview
Issue Details (click to expand)WARNING
Files Reviewed (5 files in this incremental pass)
Verification notes: Static review only. Update/patch idempotence remains unverified: the skill has no Fix these issues in Kilo Cloud Previous review (commit f093e8e)Status: 6 Issues Found | Recommendation: Address before merge Overview
Issue Details (click to expand)WARNING
SUGGESTION
Verification notes: Static review only; no helper was executed in this read-only session, so the Playwright preflight/CORS behavior above is reasoned from the code and Playwright semantics. Updater/patch idempotence is unverified: no Files Reviewed (7 changed files in this incremental pass)
Fix these issues in Kilo Cloud Previous review (commit 7d8bcb3)Status: 11 Issues Found | Recommendation: Address before merge Overview
Unverified: The browser helper's Playwright/CORS/WebSocket behavior was reviewed statically, not executed. Upstream provenance (commit Issue Details (click to expand)WARNING
SUGGESTION
Files Reviewed (24 files)
Reviewed by deepseek-v4.1-flash-1017 · Input: 61.3K · Output: 13.2K · Cached: 727.2K Review guidance: REVIEW.md from base branch |
Imports Landing Page Guru 0.1.1 at 0dc348e, which fixes the helper issues from the automated review. Drops the top-level license, since metadata.source.license_path covers it and the updater removes it anyway. Removes the MCP's skills reference: CONTRIBUTING.md asks for the skill and its first companion reference in separate steps, so the archive exists before the catalog points at it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
| const response = contract.form.responses[scenario === 'retry' || scenario === 'invalid' ? 'accepted' : scenario]; | ||
| return route.fulfill({ status: response.status, contentType: 'application/json', body: JSON.stringify(response.body), headers: { 'access-control-allow-origin': target.origin } }); | ||
| } | ||
| if (!['GET', 'HEAD'].includes(request.method())) { |
There was a problem hiding this comment.
WARNING: Cross-origin OPTIONS preflight is aborted and counted as an undeclared mutation
The enforced permittedOrigin now explicitly supports an endpoint on a different origin from the test page. A page-script fetch to that endpoint with a JSON body, an idempotency header, or credentials first sends an OPTIONS preflight to the same URL. That request is not POST, so it falls into this branch, matches the endpoint origin/path, increments blockedMutations, and is aborted; the configured POST never runs and the scenario fails as an undeclared mutation. The fulfilled POST response also lacks access-control-allow-methods/access-control-allow-headers, so a preflight cannot succeed even if it were forwarded. Fulfill OPTIONS for the endpoint with CORS headers (allow origin target.origin, method POST, and the requested headers) without counting it, or reject cross-origin endpoints in validation and document that only same-origin endpoints are testable.
There was a problem hiding this comment.
This one doesn't apply. With request routing on, Playwright answers CORS preflights itself on Chromium, so OPTIONS never reaches this handler. The cross-origin fixture in the upstream repo (tools/browser-check.mjs and tools/form-fixtures.mjs) sends a JSON POST with a custom header to another origin and passes all six scenarios. No change needed, and a code comment now records this.
| addCheck(report, scenario, calls === expectedCalls && blockedMutations === 0 && visible && announced && retained && identityOK, 'Expected response state, announced feedback, retained input on failure, and safe retry identity', `${calls} intercepted requests; visible=${visible}; live-region=${announced}; retained=${retained}; retry-identity=${identityOK}; undeclared form mutations=${blockedMutations}; unrelated mutations blocked=${unrelatedMutations}`); | ||
| } | ||
| } catch { | ||
| if (nativeSubmission) addCheck(report, scenario, 'not-tested', 'A submission handled by page script', 'The form navigated on submit, which this helper cannot simulate. Raw page text and values omitted.'); |
There was a problem hiding this comment.
WARNING: A missing native-validation failure is reported as not-tested
nativeSubmission is set for any navigation request to the endpoint, so when a form without working native validation navigates on the empty submit in the invalid scenario, the destroyed context throws and this branch reports not-tested instead of fail. The invalid check exists precisely to prove that empty submission is blocked, so "the form navigated on submit" is a failure there, not an unsupported case. Only downgrade to not-tested for scenarios other than invalid (for example if (nativeSubmission && scenario !== 'invalid')), and fail invalid whenever calls > 0.
There was a problem hiding this comment.
Fixed in 0.1.2. An empty submit that navigates now fails the invalid check, and a fixture covers it.
| const requestURL = new URL(request.url()); | ||
| if (requestURL.href === endpoint.href && request.method() === contract.form.destination.method) { | ||
| calls++; | ||
| if (request.isNavigationRequest()) nativeSubmission = true; |
There was a problem hiding this comment.
SUGGESTION: isNavigationRequest() also matches subframe navigations
A form that posts into a hidden <iframe> (a legacy async-submit pattern) produces a navigation request for the child frame. nativeSubmission is then set even though the parent page remains and its status could still be inspected, so every later failure is downgraded to not-tested and a real defect is hidden. Restrict the flag to top-level navigations, for example request.isNavigationRequest() && !request.frame().parentFrame().
There was a problem hiding this comment.
Fixed in 0.1.2. Only top-level navigations count as native submissions, so a hidden-iframe post now reports a real failure. It also handles a navigation that opens a new tab, where request.frame() can throw.
| const adapter = contract.verification.browser; | ||
| const endpoint = safeURL(contract.form.destination.endpoint, target); | ||
| // Values this short could match an unrelated request body by chance. | ||
| const formValues = Object.values(adapter.values).filter(value => typeof value === 'string' && value.length > 3).flatMap(value => [value, encodeURIComponent(value)]); |
There was a problem hiding this comment.
SUGGESTION: The form-value heuristic under- and over-matches
encodeURIComponent does not encode spaces as +, so an application/x-www-form-urlencoded body (URLSearchParams, common with fetch) encodes John Smith as John+Smith and is not recognized as carrying form values; boolean checkbox values are excluded entirely. Conversely, any 4+ character value such as test matches an unrelated body by substring and fails the scenario. Parse the body by content type and compare decoded field values (or at least add the URLSearchParams encoding) instead of raw substring matching.
There was a problem hiding this comment.
Partly fixed in 0.1.2. Bodies that encode spaces as + are now matched. Booleans and short values stay out on purpose, because matching true would flag unrelated posts. The docs now ask for distinctive synthetic values. A false match fails the test rather than hiding a leak, which is the safer direction.
| const opaque = value => typeof value === 'string' && /^[a-zA-Z0-9_-]{1,100}$/.test(value); | ||
| const personal = value => [ | ||
| /@|bearer\s|https?:\/\/|(?:\+?\d[\s().-]*){9,}/i, | ||
| /\b\d{1,3}(?:\.\d{1,3}){3}\b|(?:[0-9a-f]{0,4}:){3,}[0-9a-f]{0,4}/i, // IP addresses |
There was a problem hiding this comment.
SUGGESTION: The IPv6 pattern both misses and over-matches
(?:[0-9a-f]{0,4}:){3,} requires at least three colon-terminated groups, so common compressed addresses such as ::1, fe80::1, and fd00::1 do not match. It does match any 3+ colon hex-like value such as 12:34:56:78, and the IPv4 branch matches version-like 1.2.3.4 without octet range checks, so non-personal values fail property-allowlist. Use an explicit IPv6 matcher that handles :: and bound IPv4 octets to 0-255, or scope the check to IP-named properties.
There was a problem hiding this comment.
Fixed in 0.1.2 with Node's net.isIP, which catches ::1 and fe80::1 and no longer flags 12:34:56:78 or 999.1.2.3. A four-part version like 1.2.3.4 is still flagged, because it is also a valid IP address.
| const personal = value => [ | ||
| /@|bearer\s|https?:\/\/|(?:\+?\d[\s().-]*){9,}/i, | ||
| /\b\d{1,3}(?:\.\d{1,3}){3}\b|(?:[0-9a-f]{0,4}:){3,}[0-9a-f]{0,4}/i, // IP addresses | ||
| /\b(?:sk|pk|rk|sess|session|api_?key)[_-][a-z0-9]{4,}|\b[a-z0-9]{32,}\b/i // keys and session IDs |
There was a problem hiding this comment.
SUGGESTION: Case-insensitive short prefixes flag ordinary values
With /i, \b(?:sk|pk|rk)[_-][a-z0-9]{4,} matches ordinary SKU-like values such as SK-2024 or PK-1000, so a benign allowlisted property value fails property-allowlist. The long-token branch also cannot match a 32+ character token glued to a _ prefix because _ is a word character. Make the short prefixes case-sensitive or format-specific (sk_live_, pk_test_), and use non-word-character lookarounds for the long-token branch.
There was a problem hiding this comment.
Fixed in 0.1.2. Provider prefixes are now case-sensitive, so SK-2024 passes, and long tokens are caught when joined to a prefix.
Imports Landing Page Guru at 1b52e96, which fixes the helper issues from the second automated review, including a regression from 0.1.1 where a form that posts on an empty submit was reported as not-tested. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
| /@|bearer\s|https?:\/\/|(?:\+?\d[\s().-]*){9,}/i, | ||
| /\b(?:sk|pk|rk)[_-][a-zA-Z0-9]{4,}/, // provider keys; case-sensitive so SK-2024 stays a SKU | ||
| /\b(?:sess|session|api_?key)[_-][a-z0-9]{4,}|[a-z0-9]{32,}/i // session IDs and long tokens | ||
| ].some(pattern => pattern.test(String(value))) || String(value).split(/[\s,;=]+/).some(part => isIP(part)); |
There was a problem hiding this comment.
WARNING: IP detection misses addresses embedded in a larger token
isIP(part) only succeeds when an entire whitespace/comma/semicolon/= token is a bare address. The previous word-boundary IPv4 pattern matched an address inside 10.0.0.1:8080, (10.0.0.1), or 10.0.0.1.; those tokens now fail isIP, so a property carrying an IP passes property-allowlist. Strip surrounding punctuation and brackets, then retry the address and its host portion before a :port (for example isIP(part) || isIP(bare) || isIP(bare.split(':')[0])).
Imports Landing Page Guru at 27927b9, which fixes the warning from the third automated review: an IP address next to a port, brackets, a zone or a final full stop (10.0.0.1:8080, (10.0.0.1), 10.0.0.1., [::1]:8080, fe80::1%eth0) now fails property-allowlist. Values that only look like addresses (12:34:56:78, 999.1.2.3, 1.2.3.4.5) still pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
| description: Build, audit, or improve landing pages with clear offers, persuasive evidence-backed copy, distinctive responsive design, working forms, and meaningful measurement. Use for product waitlists, B2B demo pages, local-service inquiries, lead magnets, landing-page conversion reviews, and repairs to a page's lead-capture or success-event flow. Preserve the user's existing stack and brand; adapt to available tools. | ||
| metadata: | ||
| author: Pageree | ||
| version: "0.1.3" |
There was a problem hiding this comment.
SUGGESTION: toolVersion in report metadata still reports 0.1.2
scripts/lib/common.mjs (reportFor) hardcodes toolVersion: '0.1.2', and it was bumped in lockstep with metadata.version in 0.1.1 and 0.1.2 (see the earlier commits in this PR). This release raises the skill to 0.1.3 without touching it, so validate-events.mjs and test-form-flow.mjs reports produced by the 0.1.3 package identify themselves as 0.1.2. For a skill whose output is meant to be evidence, that breaks the report's provenance. Bump it to 0.1.3 upstream (ideally derive it from the package version) and re-import.
Helper reports now carry the skill's own version (the review's open suggestion: 0.1.3 still reported toolVersion 0.1.2), and the Pageree handoff discloses a commercial service with a limited free plan and paid plans. Pageree has offered a free plan since 2026-11-15, so the MCP prerequisite no longer says a trial or paid plan is required. Source: Pageree/landing-page-guru-skill at 32ddf8b.
Adds Pageree as a remote MCP server, and its open-source Landing Page Guru skill as a standalone skill.
Pageree publishes landing pages with working lead forms from the assistant, then reads visitor analytics section by section to propose page changes.
skills:to the MCP once Package Skills has published the archive.I'm the founder of Pageree. It's a paid product with a 30-day free trial.
🤖 Generated with Claude Code