Repository navigation
Tags
Tags
Report the live deployment from the Express and Node guards too (#485) Only the fetch-style guard told Patchstack it was serving live traffic. A guard mounted as Express or Node middleware never did, so an Express app published on a host that does not label its builds stayed "not deployed" on the dashboard. Co-authored-by: Claude <[email protected]>
Wire the server entry, not a test file beside it, and name runtime re… …ports in the install prompt (#483) protect skipped test directories but not a test kept beside the code it covers, so entry.test.ts could be chosen over entry.ts: the test was wired and the real entry was reported as a second, unguarded server. The install prompt now lists the reports the guard sends for requests its rules match or block, and says "reviewed" instead of "vetted". AGENT-INSTALL.md says the runtime check needs a URL the CLI can reach, which in a hosted builder is the preview's public address. Co-authored-by: Claude <[email protected]>
Install with the package manager first, and accept the reporting opt-… …outs (#481) The install prompt now names the package as Patchstack's package for JavaScript and Node.js apps, separate from its WordPress plugin, and asks for an ordinary install with the project's package manager before the CLI runs. The command drops --yes: with the package installed, npx runs the installed copy. AGENT-INSTALL.md adds the fallback for a sandbox that will not run npx: node node_modules/@patchstack/connect/dist/cli.js. protect --check now accepts a guard whose only change is a documented reporting opt-out (reportDetections: false, reportFirewallLog: false). Any other edit still makes it a custom guard. The docs point to the environment variables first. A new synthetic persona, site-builder, covers the three pressures together. Co-authored-by: Claude Opus 5.5 <[email protected]>
Ask a hosted builder's assistant to check setup's files were kept (#480) In a Lovable or Replit project, setup's output now asks the assistant, just before its last line, to run doctor before it replies, and to tell the person setup did not finish if doctor says Patchstack is not installed. The files setup writes reach the project only if the builder keeps them, and the assistant's reply is all the person sees. Co-authored-by: Claude Opus 5.5 <[email protected]>
Finish setup with the newest version a project accepts (#467) pnpm 11 and later skip a package version published in the last day, so the release `npx` just fetched is often refused when setup adds itself. Setup then stopped before writing anything, and the assistant running it could still report success. When the exact version is refused, setup now lets the package manager add the newest version it allows and hands over to that version's own setup with the same arguments, so the files written match the installed package. If that fails too, the output ends on a plain line saying setup did not finish. Both are disclosed in README.md and AGENT-INSTALL.md. Co-authored-by: Claude Opus 5.5 <[email protected]>
[NO-TICKET] Restore a hosted builder's lost key with `login`, in the … …CLI's messages and the docs (#465) * Tell agents in a hosted builder to recover a lost key with `login` A hosted builder's workspace may not keep the git-ignored key file, so a later scan finds no key. AGENT-INSTALL.md now says to run `login` there, never to ask the person for a key or open the builder's secrets form, to remove a wrong PATCHSTACK_API_KEY secret first, and to scan and publish again after approval because rotating retires the read-only credential the published app carries. Co-Authored-By: Claude Opus 5.5 <[email protected]> * Restore a hosted builder's lost key with login, in the CLI's own messages In a Lovable or Replit project, the messages for a missing or refused key now say to run login in the builder's workspace, then scan and publish again. They no longer offer PATCHSTACK_API_KEY, which sent agents to the builder's secrets form to ask the person for a key they were never shown. login's own output says the same there, and on any machine warns when PATCHSTACK_API_KEY or PATCHSTACK_PULSE_AUTH is set, since either one is read before the key login saves. A builder's project deployed through a platform that identifies itself keeps the environment-variable advice. Co-Authored-By: Claude Opus 5.5 <[email protected]> * Say the lost-key steps in plain words The builder messages no longer say "read-only credential", "git-ignored" or "read before": the agent passes these on to the person, so they say what happens instead (the live site keeps getting new protection rules, the secret overrides the new key). Co-Authored-By: Claude Opus 5.5 <[email protected]> --------- Co-authored-by: Claude Opus 5.5 <[email protected]>
[NO-TICKET] Mark a one-line root layout, and end setup's output with … …the claim link (#461) * Add the production marker to a root layout written on one line The marker step looked for <head> or <body> only at the start of a line. A v0 app's Next.js layout is written on one line (`return <html lang="en"><body …>…`), so it found no anchor, added no marker, and the published page never said which build it ran: the dashboard could not identify the live build. It now also anchors on a tag in the chain a `return` or an arrow function opens with, after trying line starts as before. A tag inside a string is not in such a chain and is still ignored. Co-Authored-By: Claude Opus 5.5 <[email protected]> * End setup's output with the line to pass on, and the claim link An assistant that ran setup reported only that the command succeeded and left out the summary, so the person never got the link to connect their Patchstack account. The instruction to pass the summary on came before it, in the middle of long output that an assistant may cut or skim. Setup's output, and summary's, now end with one line, the one an assistant is surest to read: include the summary above as written, and, while the site has no owner, with this link to connect the account. AGENT-INSTALL.md says so. Co-Authored-By: Claude Opus 5.5 <[email protected]> --------- Co-authored-by: Claude Opus 5.5 <[email protected]>
Refresh rules on Vercel without waiting out a backoff (#459) A Next.js middleware on Vercel calls the guard with the request alone, so the guard found no waitUntil and its 30-second background rules check was cut off when the response went out. Each cut-off counted as a failed refresh and pushed back the refresh a request waits for, up to 8 times the 5-minute interval, so a copy of the app could keep a deleted rule for about 40 minutes. - A background check that fails or is cut off now backs off only its own deadline; the waited refresh still comes due on time. - hostWaitUntil also reads Vercel's request context (globalThis[Symbol.for('@vercel/request-context')].get().waitUntil), the place @vercel/functions reads it, so background checks on Vercel finish after the response. Co-authored-by: Claude Opus 5.5 <[email protected]>
Give the agent one recipe for wiring any Node.js app (#448) Setup wires the frameworks it recognises and leaves a generic guard for everything else. The guide described the procedure in general terms, and agents filled the gap by writing their own createProtection guard, which has no read-only rules credential and cannot fetch rule updates. AGENT-INSTALL.md now has a short recipe under "Completing guard wiring": use the guard file setup wrote (never write your own or copy the .patchstack cache), wrap the app's one request entry with the export that fits its shape (Fetch handler, Node (req, res), Express/Connect, Fastify), run setup again so the guard's rules file gets the site's rules and the credential, then check with protect --check and --runtime. Also corrects the refresh interval it states: 5 minutes, plus a background re-check every 30 seconds while serving, not 45 minutes. Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>
Let the deployed guard name the build it runs (#456) The guard's live report sends `_patchstack.manifest_checksum` from its rules file, but nothing wrote that field, so a live report could only say that some build was serving. The `prebuild` scan now records the checksum of the package list being built there, beside the map stamp, before the bundler reads the file. A server-rendered deployment with no marker on its pages can then say which build it is. Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>
PreviousNext