DEV Community: kyisaiah47

KitGrade: SaaS starter kits scored with visible evidence

kyisaiah47 — Sun, 11 Oct 2026 10:31:54 +0000

kitgrade.thecompound.tech

KitGrade scores SaaS starter kits against evidence buyers can check, using a Next.js application and frozen JSON data. Developers use it to compare what a kit includes, how it is maintained, and which claims have public support.

The problem

A starter kit can advertise authentication, billing, documentation, and support without giving a buyer one place to verify those claims. KitGrade checks the repository, dependency manifest, documentation, pricing page, release history, and public support channel.

The distinction between "We ran it" and "Read, not run" matters. A build transcript records commands, exit statuses, and durations, while a documented result records what the repository or vendor page exposes.

How it works

KitGrade keeps a hundred-point scorecard divided across maintenance, completeness, transparency, documentation, and support. A capability that cannot be observed publicly remains unmeasurable instead of becoming a zero.

The capture script reads active kit rows, scores, and hands-on runs from the KitGrade engine database.
It joins each kit to its score and run transcript, then checks that the component figures add back to the stored total.
It writes stats.json, method.json, kits.json, and kit-detail.json into the repository.
The Next.js data layer reads those files through src/lib/data.ts, while src/lib/derive.ts turns stored breakdowns into score spans, evidence labels, and capability rows.
The console renders each score beside the evidence source, the measurement state, and any recorded install transcript.

The repository check fails when a kit total does not reconcile, a component exceeds its maximum, or the method maxima do not add to one hundred. The capture also refuses to write a set with inconsistent figures.

A worked example

KitGrade records Bullet Train as a Rails, Ruby, PostgreSQL, Stripe, and Hotwire kit with MIT licensing. Its stored score is 92.2, with 30 points for maintenance, 23.2 for completeness, 20 for transparency, 13 for documentation, and 6 for support.

The hands-on transcript records a clean clone, a successful frozen Yarn install, and a failed build because configuration is still required before the application runs.

How it compares

ShipFast presents a Next.js boilerplate for SaaS and AI applications and sells access through its own site. Pick ShipFast when you want a commercial Next.js starting point with the vendor's included feature set.

Bullet Train describes itself as an MIT-licensed Rails-based framework with authentication, teams, permissions, billing, and other application features. Pick Bullet Train when your application belongs in Rails and you want an open-source starting point.

SaaS Pegasus is a Python and Django SaaS boilerplate with flat pricing and no recurring subscription fees. Pick SaaS Pegasus when Django is the deciding constraint and a one-time license fits the project.

KitGrade differs by recording the evidence behind each comparison. Its completeness score checks the repository and file tree for open-source kits, while its transparency score includes public source, licensing, pricing, release history, and documentation.

Limits

KitGrade does not replace reading the source repository or testing a kit in the application you plan to build. Its score only covers the five published components and the evidence those components can observe.

A missing public surface is recorded as unmeasurable when the method cannot inspect it. That state is separate from a measured zero.

Try it

Start with a kit from the bench, then open its scorecard and read the evidence attached to each component. The Bullet Train record is a useful first example because it includes both repository evidence and a hands-on install transcript.

Four consumer tools: ScriptProbe, ListRun, MatchLine and BreachProbe

kyisaiah47 — Fri, 09 Oct 2026 09:44:25 +0000

This issue carries consumer tools for checking package install scripts, directory submissions, resume evidence, and app security. The outputs are a verdict, a run sheet, a tailored PDF, and a security report.

This issue

These products turn narrow tasks into concrete results you can inspect.

ScriptProbe


	ScriptProbe checks an npm package before you install it. It reads the tarball and lifecycle scripts as text, matches fixed patterns, and reports install scripts, network access, native builds, child processes, and publishing-account changes without executing the package. For a JavaScript developer or maintainer, the free check returns a verdict and the matching source text. You pay $15 once for a 30-day watch that rechecks a package.json or repository daily and emails changes.

ScriptProbe checks an npm package before you install it. It reads the tarball and lifecycle scripts as text, matches fixed patterns, and reports install scripts, network access, native builds, child processes, and publishing-account changes without executing the package. For a JavaScript developer or maintainer, the free check returns a verdict and the matching source text. You pay $15 once for a 30-day watch that rechecks a package.json or repository daily and emails changes.

ListRun


	ListRun is a directory-submission run sheet for a product owner who wants to see where a listing qualifies and what each submission requires. You answer what the product is, what it plugs into, and what it was built with, then ListRun shows the queue, the directories it will not attempt, and the order it reaches them. Each row carries a directory, submission method, qualification state, and note, with a receipt for every one. You pay $9 once for the Launch run or $19 once for the Full list.

ListRun is a directory-submission run sheet for a product owner who wants to see where a listing qualifies and what each submission requires. You answer what the product is, what it plugs into, and what it was built with, then ListRun shows the queue, the directories it will not attempt, and the order it reaches them. Each row carries a directory, submission method, qualification state, and note, with a receipt for every one. You pay $9 once for the Launch run or $19 once for the Full list.

MatchLine


	MatchLine checks a resume against a job posting, line by line. It reads every stated requirement, points to the resume line that proves it, and separates gaps with a line to add from gaps with nothing to build on. For a job seeker, the free check returns the reading and both documents without keeping an archive. You pay $5 for a tailored PDF.

BreachProbe


	BreachProbe checks whether a shipped app keeps one signed-in user's rows away from another. You paste an app URL and get a security score with checks for exposed keys, anonymous table access, headers, auth flows, and row-level isolation. For an app builder shipping user-owned data, the report uses two test users and compares table access rather than reading policy text. The result is a security report with findings, comparisons, and suggested fixes.

BreachProbe checks whether a shipped app keeps one signed-in user's rows away from another. You paste an app URL and get a security score with checks for exposed keys, anonymous table access, headers, auth flows, and row-level isolation. For an app builder shipping user-owned data, the report uses two test users and compares table access rather than reading policy text. The result is a security report with findings, comparisons, and suggested fixes.

One issue a week, and one shipped product taken apart every month. What each product does, what it cost to build, what the pipeline behind it looks like, and what the numbers did, read off the repository and the live site, not written from memory. Join the list.

OpenCal puts shared free time on one link

kyisaiah47 — Wed, 07 Oct 2026 17:40:26 +0000

One link for open time

OpenCal shows open times from connected calendars on one link. Visitors can view available times up to 90 days ahead. Visitors can switch between Week, 2 weeks, and Month views. Visitors see the account name and selected profile details. Visitors never see event titles, places, or guests. OpenCal reads busy times instead of event details.

OpenCal is free to use. You get a link by signing in with Google or Microsoft and selecting Get your link. Settings lets you change the link handle and choose which calendars count as busy.

Maya sends one link

Maya is a designer at Northwind. Sam is a recruiter at Brightpath Talent. Sam needs a 30-minute call with Maya during the current week or the next week.

Before OpenCal, Sam emails Maya for available times. Maya checks Google Calendar, the company Teams calendar, and iCloud. Maya types free times into the reply. Maya checks Teams again after finding one incorrect time.

With OpenCal, Maya connects Google and iCloud. Maya signs in with Google and connects iCloud with an Apple Account. OpenCal reads only busy times from those calendars. Event titles stay in the connected calendars.

Maya sends one OpenCal link to Sam. The link shows free time across the connected calendars.

Teams from the screen

Most companies block outside apps from reading Teams calendars. OpenCal reads a company Teams calendar from the screen when a direct calendar connection cannot work.

Maya opens the Teams week view. Maya presses Read from screen and selects the Teams window. OpenCal reads one frame in the browser. The screenshot never leaves the browser. OpenCal saves only busy times and stores those times encrypted. Maya checks the times before OpenCal saves them.

OpenCal can also read an Outlook desktop calendar from the screen. The Chrome and Edge extension reads Outlook on the web while the account stays signed in. A private .ics feed can provide another calendar source.

Times for both people

Sam opens Maya's link and signs in with Google or Microsoft. OpenCal compares Maya's calendars with Sam's calendar. OpenCal sends Sam only times that work for both people.

Sam books Thursday at 11:00. OpenCal saves the booking for Maya and Sam. OpenCal keeps the booking in OpenCal by default. Maya and Sam can turn on calendar booking in Settings under Booking. Google or Outlook then adds the meeting to the selected calendar and sends the invite.

The link lets anyone book when the owner selects Anyone can book. The owner can create one-time links that allow one booking. The owner can also turn booking off while keeping the free times visible. A visitor must sign in before booking.

Calendars that connect

OpenCal connects several Google accounts, personal and work Microsoft accounts, and an iCloud account. OpenCal can read Teams and Outlook calendars from the desktop screen. The browser extension can read Outlook on the web. OpenCal can read any calendar with a private .ics address.

OpenCal checks connected calendars when a page or link opens. OpenCal subtracts busy times from the hours and days selected in Settings. OpenCal ignores events marked free, canceled, or declined. OpenCal counts tentative events as busy.

OpenCal gives each person one link for shared availability and booking. Get a free link at:

https://opencal.link

8bf407d made FAQ spacing part of the verification contract

kyisaiah47 — Fri, 02 Oct 2026 23:05:46 +0000

The defect

Commit 8bf407d was titled "Fix Simple help spacing and assert internal section gaps".

The FAQ check opened "What does BreachProbe do?" but did not inspect the support section's internal spacing.

The insight

The fix gave that layout a measurable contract. The paragraph gets 16px of top padding. The links sit 24 pixels below the paragraph.

BreachProbe is where this case lives: the verification now checks the spacing before continuing with the FAQ interaction.

The code

.bp-section { border-top: 1px solid var(--rule); padding: 48px 0; scroll-margin-top: 24px; }
.bp-support > h2, .bp-support > p { margin: 0; }
.bp-support > p { padding-top: 16px; }
.bp-support > .bp-home-links { margin-top: 24px; }
.bp-section-intro { display: grid; grid-template-columns: 1fr 1fr; align-items: end; gap: 6vw; padding-bottom: 28px; }
.bp-section-intro h2, .bp-result > h2 { font: 500 32px/1.25 var(--sans); letter-spacing: -.035em; padding-top: 14px; }

The test reads the computed padding and the distance between the paragraph and link group, then asserts 16px and 24.

The takeaway

I changed the FAQ check from opening one control to asserting the support spacing first.

The record: commit 8bf407d in BreachProbe, dated 2026-10-01.

Check your own case on BreachProbe.

Four compliance filings with a document for each one

kyisaiah47 — Thu, 01 Oct 2026 22:58:29 +0000

Four compliance filings, each with a document behind it. This issue covers tools that turn a compliance question into a report, a clock, a registry finding, or a repair binder.

What this issue covers

These pages turn compliance requirements into documents and recorded findings for the people responsible for acting on them.

PartsProof


	PartsProof helps makers check whether Article 14 reporting applies to products sold or shipped into the EU. Its scope checker asks six questions and returns an in-scope, out-of-scope, or edge-case answer with the obligation dates. The site also generates a software bill of materials from any public GitHub repository. The result is a security document that can be produced without an account.

PolicyDrift


	PolicyDrift loads a site, lists every third-party tracker it contacts, and compares those findings with the privacy policy linked from the page. It marks trackers the policy does not mention. A site owner gets a registry of vendors and findings such as undisclosed trackers, premature claims, overclaimed descriptions, and pages with no policy to read.

TrustDesk


	TrustDesk gives platforms a report form for the federal 48-hour intimate-image removal law. Each report gets a 48-hour clock and a log recording what the platform did and when. The desk also checks whether a request contains the required signature, image location, good-faith statement, and contact information. Platform operators get the request record, the deadline, and the action history together.

TrustDesk gives platforms a report form for the federal 48-hour intimate-image removal law. Each report gets a 48-hour clock and a log recording what the platform did and when. The desk also checks whether a request contains the required signature, image location, good-faith statement, and contact information. Platform operators get the request record, the deadline, and the action history together.

CivicBinder


	CivicBinder checks public entity websites for accessibility problems and turns the findings into a repair binder. Its index tracks public sites and shows when each site was last scanned. A city clerk or accessibility lead gets a report of the findings, the relevant deadline information, and a list of repairs to work through.

Four compliance filings with a document behind each one

kyisaiah47 — Thu, 01 Oct 2026 13:53:36 +0000

Four compliance filings share a document trail from intake to the artifact someone needs to act. These pages turn scope answers, tracker findings, removal notices and accessibility checks into records you can use.

What these pages cover

These pages cover compliance work where the output is a policy, report, runbook, notice or repair record.

PartsProof


	PartsProof checks whether a software or connected hardware product falls within the EU Cyber Resilience Act. Its scope checker asks six questions and returns the scope result, obligation dates and supporting documents. Makers get a disclosure policy, a 24-hour reporting runbook, a software bill of materials and a checklist for 2027. The public repository SBOM generator reads dependency manifests and produces a CycloneDX document.

PartsProof checks whether a software or connected hardware product falls within the EU Cyber Resilience Act. Its scope checker asks six questions and returns the scope result, obligation dates and supporting documents. Makers get a disclosure policy, a 24-hour reporting runbook, a software bill of materials and a checklist for 2027. The public repository SBOM generator reads dependency manifests and produces a CycloneDX document.

PolicyDrift


	PolicyDrift checks a website against the privacy policy it links to. It lists the third parties the page contacts, then marks the trackers the policy leaves out. Site owners get a report showing the gap between page behavior and policy text. The check is free, and the report records the findings by site.

TrustDesk


	TrustDesk gives platforms a removal-request route for nonconsensual intimate imagery. A reporter submits the URL, statement and electronic signature, then receives a tracking code and a 48-hour deadline. Operators get an append-only record and an Evidence Export in PDF. TrustDesk also generates a TAKE IT DOWN Act notice from a platform name, site URL and monitored removal address.

CivicBinder


	CivicBinder checks public entity websites for accessibility failures and places each result in a public record. The record shows a grade, the failure level and the Title II dates displayed for the site. City and county staff get an accessibility report they can use to see what needs attention before the listed deadlines. The page also provides a deadline guide and a request path for a free report.

CivicBinder checks public entity websites for accessibility failures and places each result in a public record. The record shows a grade, the failure level and the Title II dates displayed for the site. City and county staff get an accessibility report they can use to see what needs attention before the listed deadlines. The page also provides a deadline guide and a request path for a free report.

Four platform tools for builders

kyisaiah47 — Wed, 30 Sep 2026 21:24:51 +0000

This issue carries four platform tools with a visible surface for comparing, rendering, measuring, or submitting work.

A short field guide

Each section names what the product does, who it serves, and what it gives back. The pages keep the work concrete, from dated vendor rates to frozen launch rankings.

StackTab


	StackTab prices a database, auth service, payment provider, and host together. A developer comparing infrastructure gets monthly totals at 1,000, 10,000, and 100,000 users, with each rate dated from the vendor's pricing page. The console shows the selected services and their combined bill. Its catalogue also records plans, rates, and which vendor claims have drifted.

Tearline


	Tearline is a custom element that renders HTML as a thermal receipt and exports the result as a PNG. A frontend developer gets a DOM component with six attributes, three methods, and no dependencies. The export runs through four named stages, from flattening the shadow tree to encoding the canvas as a PNG blob. Text and CSS are inlined automatically, while images must already be data URIs.

Toolproof


	Toolproof is a register for AI agent tooling that reads indexes from source and records what they find. A researcher or builder gets dated measurements, methodology notes, badges, and an open read API. The register keeps each index's own name and method, then shows whether it answered or fell behind. Its pages attach figures to the index and capture date that produced them.

ShipWall


	ShipWall is a launch board for products built with tools such as Lovable, v0, Bolt, Cursor, Claude Code, or Replit. A builder submits what they made, which tools they used, and how much they edited by hand. The board has ten launch slots per day, assigns a date after review, and freezes the ranking at midnight UTC. Each frozen row keeps its launch date, rank, votes, build method, and hand-editing note.

ShipWall is a launch board for products built with tools such as Lovable, v0, Bolt, Cursor, Claude Code, or Replit. A builder submits what they made, which tools they used, and how much they edited by hand. The board has ten launch slots per day, assigns a date after review, and freezes the ranking at midnight UTC. Each frozen row keeps its launch date, rank, votes, build method, and hand-editing note.

CivicBinder scans public websites and turns findings into a repair binder

kyisaiah47 — Tue, 29 Sep 2026 20:47:17 +0000

civicbinder.org

CivicBinder's fulfillment lane stopped after repeated network failures because Node was choosing DNS addresses in an unstable order. Commit ad21608 fixed the failure by setting IPv4-first resolution before the helper's network calls. The live CivicBinder page shows what that work produces: a scan, a document inventory, policy and statement text, and a dated remediation plan.

CivicBinder

Accessibility evaluation

The main page describes an automated WCAG 2.1 AA evaluation covering up to 150 pages and 100 PDFs, with findings ranked by severity.

The evaluation turns scan results into findings that can be explained and handed to the person responsible for repairs.

Document inventory

The binder outline includes a document inventory covering every PDF on the site and its accessibility status.

The inventory keeps documents inside the same dated scope as the page scan, so the binder records what was checked rather than only describing the website generally.

png)

Accessibility Policy

The live binder outline includes an Accessibility Policy prepared for adoption by the governing body.

The policy gives the entity a document it can take through its regular meeting process instead of leaving the scan as a standalone technical report.

Accessibility Statement

The binder includes an Accessibility Statement ready to publish, with a feedback mechanism.

The statement turns the findings into public-facing text and gives visitors a route for reporting accessibility problems.

png)

Remediation plan

The binder outline includes a dated, phased remediation plan described as vendor-executable.

The plan orders repairs by severity and gives the work a schedule tied to the scan and the entity's applicable Title II date.

Pages evaluated

The binder includes an appendix listing the pages and documents in scope, along with the scan date.

That appendix limits the claim to the material CivicBinder actually evaluated, which makes the resulting record specific to the site state captured by the scan.

Fulfillment network helper

The commit named ad21608 changed ops/fulfill/services.mjs after the fulfillment lane exited with the same failure code repeatedly. The commit message identifies unstable DNS address ordering as the cause.

 import { readFileSync, writeFileSync, existsSync, chmodSync } from 'node:fs';
 import { execFileSync } from 'node:child_process';
+import { setDefaultResultOrder } from 'node:dns';
 import { join, dirname } from 'node:path';
 import { fileURLToPath } from 'node:url';

export const HERE = dirname(fileURLToPath(import.meta.url));
+setDefaultResultOrder('ipv4first');
 const ENV_CACHE = join(HERE, '.env');

The fix sets Node's default DNS result order before the helper reaches its network calls. I kept the change at the helper boundary, where the failure occurred, and left the fulfillment operations around it unchanged.

Four AI agents that leave a human approval step

kyisaiah47 — Mon, 28 Sep 2026 20:31:54 +0000

This issue follows four AI agents that leave a human approval step before a write, retry, reply, or send. Their pages show what each agent reads, what it produces, and where a person can still stop the action.

Approval in the loop

These are short field notes on the queues, documents, drafts, and receipts each agent produces.

TriageDesk


	TriageDesk reads a shared support inbox overnight, classifies each message, and drafts one reply per thread. It is for teams working from Microsoft 365 or Gmail who want a morning queue instead of a wall of unread mail. The queue shows the source message beside the draft, with options to approve, edit, or kill it. Complaints, legal threats, and cancellations go to a human undrafted, while approved replies sit for 120 seconds before dispatch.

TriageDesk reads a shared support inbox overnight, classifies each message, and drafts one reply per thread. It is for teams working from Microsoft 365 or Gmail who want a morning queue instead of a wall of unread mail. The queue shows the source message beside the draft, with options to approve, edit, or kill it. Complaints, legal threats, and cancellations go to a human undrafted, while approved replies sit for 120 seconds before dispatch.

MatchRail


	MatchRail runs a three-way match across bills, purchase orders, and receipts for a month-end close. It is for bookkeepers who need to see what disagrees before a payment run. Each disputed row shows the documents and the reason, such as a missing purchase order or a second copy of an invoice. An approved correction waits 60 seconds, writes an audit row, and re-reads the documents before writing to the ledger.

MatchRail runs a three-way match across bills, purchase orders, and receipts for a month-end close. It is for bookkeepers who need to see what disagrees before a payment run. Each disputed row shows the documents and the reason, such as a missing purchase order or a second copy of an invoice. An approved correction waits 60 seconds, writes an audit row, and re-reads the documents before writing to the ledger.

LeadGrade


	LeadGrade scores every inbound lead overnight and ranks the queue highest-first. It is for sales teams using HubSpot who need the score and its reason beside each lead. Every lead starts at 20 and can add points from form intent, role, company size, email domain, industry, buying language, and completed fields. Approve holds the HubSpot write for 60 seconds, while Kill writes nothing.

CardChase


	CardChase recovers failed subscription charges by reading decline codes, staging retries on a ladder, and stopping hard declines. It is for subscription businesses that need a queue sorted by recurring charge at risk. Each row carries the decline code, issuer advice, attempts spent, and the next rung. A retry can be approved into a 30-second kill window, while a hard decline or a charge at its four-retry cap is stopped.

CardChase recovers failed subscription charges by reading decline codes, staging retries on a ladder, and stopping hard declines. It is for subscription businesses that need a queue sorted by recurring charge at risk. Each row carries the decline code, issuer advice, attempts spent, and the next rung. A retry can be approved into a 30-second kill window, while a hard decline or a charge at its four-retry cap is stopped.

CiteRank tracks whether answer engines cite your site

kyisaiah47 — Sun, 27 Sep 2026 20:13:10 +0000

citerank.thecompound.tech

CiteRank's commit 9435a68 fixed a layout defect in src/app/globals.css: prose width rules were leaving dead space beside the console rail. The fix removed those limits so content could fill its column.

CiteRank

The board

The board shows one row per buyer question and one column per answer engine. Each result distinguishes cited, named, and absent.

The board tracks whether an engine links your domain, names your brand without a link, or leaves you out. Its status line shows the last recorded answer date.

Free check

The Free check view presents a question field and a check control for a single prompt. It gives the console a direct way to inspect one answer before tracking it.

The control runs a one-off check and reports whether the answer cites, names, or omits the supplied site. CiteRank keeps that result separate from its weekly board.

llms.txt

The llms.txt view explains the file format and shows the content that can be provided to answer engines. The page presents the document as a readable text artifact.

The view helps inspect whether a site has a clear machine-readable description. It keeps the document itself visible instead of reducing the check to a score.

AEO tools

The AEO tools view presents structured tables for improving how a site appears in answer-engine results. Its controls organize the work around questions, sources, and answer visibility.

The page turns a missing citation into a concrete research task by showing the answer context around the result. It keeps cited, named, and absent states distinct.

How to rank

The How to rank view contains guidance for improving visibility in ChatGPT and related answer engines. Its sections explain how the board's result states differ.

The page connects a result to the source material an engine used. It makes the cited-source list part of the workflow instead of treating visibility as a single number.

The width fix

The defect lived in three selectors that applied the reading measure to content that already had a console column.

- .doc > section > p.doc > section > ul,
- .doc > section > ol.doc > section > h3 {
-   max-width: var(--measure);
- }
+ .doc > section > p.doc > section > ul,
+ .doc > section > ol.doc > section > h3 { }

The same change removed the measure from band headers and ledger captions. I kept the full-width rule for tables because those artifacts use the whole column.

CertScope determines which CPSC rules apply to a SKU

kyisaiah47 — Sat, 26 Sep 2026 20:07:31 +0000

certscope.thecompound.tech

CertScope starts with a product description, an HTS code, and an age group. It turns that input into a rule decision, a certificate draft, and a list of fields that still need evidence.

CertScope

SKU determination form

The main page opens with a determination form containing a product description, an HTS code field, and age-group controls for children's, general, or unknown products.

The Determine control sends those values to the scope service. CertScope returns a certificate type, matched rules, testing requirements, and the rules that need to be cited.

Determination stages

The determination view shows a vertical stage list beside the form. Each stage can display its measured runtime, including values below one millisecond.

The stages move from pending to completed while CertScope runs the determination. The result stays in the same workspace, so the rule register and certificate update from the submitted SKU.

Rule register

The rule register shows the CPSC dataset as separate rows. Each row can be opened to show its rule identifier, name, applicability state, and the basis for that state.

The filter control narrows the register to all rules, applicable rules, or rules that need screening. CertScope moves rows into their new groups when a determination changes.

Certificate sheet

The certificate sheet shows the document produced from the determination. Its sections include the certificate type, cited rules, required information, and the importer details that are still needed.

CertScope writes each applicable rule on its own citation line. Fields that only the importer can provide appear with a required marker instead of being filled with an assumption.

Certificate linter

The linter control replaces the determination form with a certificate text area and a lint action. Its result view lists the seven certificate elements as separate checks.

CertScope marks each element as present or missing and places those marks back onto the corresponding certificate sections. A missing lab result remains visible as missing evidence rather than becoming a passing check.

CardChase recovers failed subscription charges before churn

kyisaiah47 — Wed, 23 Sep 2026 12:10:45 +0000

cardchase.thecompound.tech

CardChase had a real failure mode in commit 04e6130: an admin lookup error could look like "no demo account," allowing a fabricated retry to continue toward payment or email. The fix touched three paths and made the lookup fail closed.

CardChase

The queue

The queue shows seven open charges, their monthly risk, the retry cap, the kill window, and three actions waiting for approval. Each row carries the customer, amount, decline reason, and current rung.

The queue lets me approve, edit, or stop the next retry. Approval stages both the card retry and the customer note behind a 30-second window.

The record

The record shows receipts for failed charges, drafted notes, staged retries, blocked attempts, capped retries, and recovered charges. Each receipt includes evidence such as the invoice, decline code, rung, or stop authority.

The record preserves why CardChase acted or stopped. A permanent decline writes its own receipt instead of appearing as an ordinary retry.

The ladder

The ladder shows four rungs across seven days: a silent retry after one day, then notices after three, five, and seven days. The schedule counts from the first failure, not from the previous attempt.

The ladder decides when a charge becomes due. Before staging anything, CardChase checks the issuer advice, hard-decline codes, and raw network response.

The kill window

The kill-window view shows the 30-second release period shared by the retry and the customer note. An approved action is staged, not sent immediately.

The window lets me undo the action before either side reaches the payment or mail rail. The dispatcher cannot claim the row until the release instant.

How it works

The workflow moves through five named states: failed charge, ladder, gate, approval window, and receipt. The gate can stop the charge before the ladder produces a retry.

flowchart LR
  A[Failed charge] --> B[Ladder]
  B --> C[Stop gate]
  C --> D[Approval window]
  D --> E[Receipt]

The page explains the same path in the console's own terms. CardChase reads the failed charge, checks whether the next rung is due, stages the action, and records the evidence.

The changelog

The changelog shows four shipped changes, including the hard-decline stop, the retry cap, the shared kill window, and direct billing reads. The kill-window entry says approval stages both halves of the action.

Commit 04e6130cf4ba2666c4b8ae559f40c8c99ababd0e fixed the lookup defect in src/app/_lib/dispatch-run.ts, src/app/_lib/pass-run.ts, and src/app/auth/callback/route.ts.

diff --git a/src/app/_lib/dispatch-run.ts b/src/app/_lib/dispatch-run.ts
@@
-  const { data } = adminSupabase().auth.admin.listUsers({ page: 1, perPage: 200 });
+  const { data, error } = adminSupabase().auth.admin.listUsers({ page: 1, perPage: 200 });
+  // A broken admin lookup is not evidence that the demo user is absent.
+  if (error) throw new Error(`Unable to resolve the demo account: ${error.message}`);

I changed the lookup so an error stops the pass before it can reach a real payment or mail rail.