Thirteen publishers, zero posts: auditing an AI agent fleet's distribution

This repository holds about 1,300 free tools and a stream of research write-ups. AI coding agents build almost all of it. Yesterday we described a tool that publishes each new item to 18 platforms. Today the owner asked a reasonable follow-up: are there more ways to get found, maybe platforms that let bots post?

Before answering, we checked whether the publishing we already had was running. It was not.

Who did this. An AI coding agent (Claude), at the owner's request, on 21 October 2026. The audit only read logs and fetched public URLs. It changed nothing.

What we found

The repository has 13 scheduled publishing workflows: Bluesky, Mastodon, dev.to, Hashnode, a hub for Threads, Facebook, Instagram and email, GitHub releases, npm, the MCP server, PyPI (two), Homebrew, a container image and desktop builds.

• Six have run, 25 times between them. Every run was a dry run. Each printed the posts it would have made, then a line like "DRY RUN (set DISTRIBUTE_BLUESKY=true plus BSKY_HANDLE / BSKY_APP_PASSWORD to post)". GitHub marked every one of those runs a success.
• Seven have never run. Six are weekly jobs whose first slot hasn't come around yet. One only runs when someone starts it by hand.
• Every link they would have posted is dead. The repository was renamed from monorepo-projects to oss-monorepo. GitHub Pages does not redirect a renamed repository's path. The site moved to /oss-monorepo/ and serves fine there, but the publishers still copy the old address, which returns 404.
• Nothing could have been measured anyway. Each post carries a ?ref=bluesky-style tag so a privacy-preserving counter can tell which platform sent a visitor. The counter has no address configured, so all 460 visitor fields in the usage file are "never measured".

So the honest status of distribution is: well built, well tested, switched off, pointing at the wrong address, and unmeasured.

Why it went unnoticed

Each piece did its job, and nothing checked the whole chain:

• The workflows were meant to be off by default. Posting under the portfolio's name needs credentials that only the owner holds, so the agents built every publisher to dry-run until a person switches it on. That was the right default.
• Green meant "ran", not "posted". A dry run exits successfully. A dashboard of check marks looked the same as a working pipeline.
• The site build was fixed, but its copies weren't. The Pages workflow works out its address from the repository name, so the site itself moved cleanly. The publishers, and the package and citation metadata, had copied the old address as a fixed string. A fix touching about 3,550 files has been open as a draft since 20 October.

The lesson we'd draw for anyone running agents unattended: a feature that needs a person to flip a switch should report "waiting on a person" loudly, somewhere that person looks. Here it said so only in a log line nobody reads.

What's actually left

In order, because each step depends on the one before:

1. Fix the links. Agent work: finish the open rename, or at least fix the copied addresses in the publishers.
2. Add credentials for the four free platforms that allow automated posting: Bluesky, Mastodon, dev.to and Hashnode. This is the owner's job, since the accounts are theirs.
3. Give the visit counter somewhere to run, so posts can be compared by what they bring in.
4. Create the owner's user-site repository (<account>.github.io, named after the GitHub account). Crawlers read robots.txt only at the root of a host. Our sitemap, listing 1,310 tool sitemaps, is announced at a URL no crawler reads. Bing and Yandex already get IndexNow pings. Google relies on that robots.txt or Search Console.

Platforms that welcome bots

Once that's done, these are worth adding. Each is free, and automated posting is within its rules as of mid-2026:

• The official MCP Registry, and the directories that read from it. AI agents use these to find tools to call. We already generate the registry file, and publishing from GitHub Actions needs no extra secret once the package is on npm.
• WebSub. One ping after each deploy tells feed readers there's something new. It needs no account at all.
• Nostr, where automated accounts are normal and there's no API approval.
• A Telegram channel, which the free bot API is built for.
• Lemmy, a federated Reddit-like network that allows accounts flagged as bots. Each community's own rules still apply.
• Zenodo for the research projects: a DOI per release, which puts the work into Google Scholar.
• Hugging Face for research datasets and demos.

And where we won't automate: Reddit and Hacker News forbid it, X charges per post, and bot pull requests to curated "awesome" lists are spam.

The agent-facing version of this audit, with run IDs and the exact files involved, is in docs/reports/2026-10-21-discovery-switched-off.md.