← All posts
news

Postman approved its own hook. Then it took it back.

· 5 min read · by aft3r_d0t

Duotone ink illustration of a postman's hand slipping a sealed letter under a locked door at night.

On 25 October a developer opened an issue on Postman’s tracker with a short, uncomfortable story. He’d opened the Postman desktop app on his Mac, and afterwards his coding agents started acting strangely: Codex kept routing API work through Postman skills, and MCP authentication warnings kept popping up.

He went digging and found the cause in Postman’s own log. An update step labelled auto-existing-upgrade had installed plugins into Claude Code and Codex: claude-code:installed, codex:installed. Nobody asked him.

The issue title says it plainly: Desktop auto-installs Codex/Claude plugins; installer attempts to trust its own startup hook. The evidence he pasted was the kind you can check yourself — a timestamped log line from the app, and three lines of the agent’s global config showing a marketplace entry, an enabled plugin entry, and a trusted-hook hash.

What the plugin actually did

Two things stand out, and both can be read in Postman’s public plugin repo at the commit the reporter cites.

First, a session-start hook. Every time the agent starts, clears or compacts its context, the hook injects a block of text, wrapped in <EXTREMELY_IMPORTANT> tags, telling the agent to load Postman’s skill “before responding to any non-trivial API engineering task” and follow it. Designing, implementing, testing, documenting, deploying: pretty much any API work. The text also says the agent may enter a more specific skill directly — mocking, testing, monitoring, documentation — but the mandate is the same: the route in is through Postman’s.

Second, the installer approves that hook itself. Codex keeps a record of which hooks the user has trusted. Postman’s installer has a function, approvePostmanHook, whose own comment reads “Approves only Postman’s current hook after installation”. It lists the hooks, takes the current one, and writes its hash into that trust record through Codex’s own config API, with a version check so a concurrent edit makes Codex refuse the write. The reporter is careful here: the source shows the attempt, but the record on his machine doesn’t by itself prove who wrote it.

That distinction matters more than it looks. The code is the vendor’s, visible and checkable. The trust entry was on one person’s machine, in a file only that person could see. Both halves of the story are needed to make the case, and the reporter had both.

Why the installer did it on purpose

Worth saying, because “it was a bug” would be the comfortable version and it isn’t the one the history supports. The plugin repo has a commit from 21 October, four days before the report, called auto-approve Postman’s Codex hook. Its message explains the reasoning: Codex 0.164 binds a plugin hook’s approval to a hash of every file in the plugin, so every upstream commit invalidates the approval and the agent asks again. The fix was to pin the marketplace to a release branch so the approval survives, and to write the trust entry automatically so the user isn’t interrupted by a prompt for a hook they had already installed.

Read generously, that is a product decision about friction. Read strictly, it is an installer writing a permission record on the user’s behalf, for a mechanism whose entire purpose is to ask the user. Both readings are true at once, and only the second one is interesting.

What happened next

The story hit the top of Hacker News — 1,530 points as of 12 November, against 1,027 when our digest first flagged it. On 3 November a commenter posted what read like a vendor reply, explaining that the plugins are bundled because people struggled to set them up by hand. On 5 November a member of the Postman organisation on GitHub wrote that they’d misjudged how clear the messaging was, would redesign the experience “with the right level of control”, and added: “In the meantime, we are reverting this change.” A few hours later, the account of Postman’s founder locked the issue as closed, with 69 comments under it.

So: reported 25 October, reverted 5 November.

But the revert was not the end of it, and the follow-up is the part worth reading. The plugin’s own history over the next day is a small, unusually honest story about what happens after a vendor says “we’re reverting this”:

  • On 5 November, a commit went in that did more than revert. It stopped auto-approving the Codex hook outright and removed the code that wrote the trust record, with a commit message that says the user must make that choice. It also added a check that the hook stays untrusted after every install and migration, and the README started telling users to review and trust the hook themselves.
  • Six hours later, that commit was reverted.
  • The next morning it was re-applied and released, as version 0.7.2.
  • The day after, another change made remove a complete uninstall: it now also removes the marketplace entries it registered and, importantly, deletes Codex’s approval of the session-start hook through the same config API — reading the version first, so a concurrent edit makes Codex refuse the write. The commit message is explicit that anything it can’t prove is Postman’s is left alone and printed as a note instead.

As of 12 November, the plugin’s Codex route no longer declares the hook at all, and the removal code carries the comment: the Codex route no longer declares this hook, but installs from 0.4.1 to 0.7.1 wrote its approval. Which is to say: if you were one of the people affected, the fix you need is not an update, it’s a cleanup. remove now takes the approval with it.