GitHub retracts mistaken Claude Code fork takedowns after cch signing reverse-engineering

• Developers say they fully reversed Claude Code's cch= request-signing scheme, then merged support into the open source free-code client, which means third-party clients can now generate the same billing-header hash without Anthropic's official binary reverse-engineering claim implementation links.
• The reverse-engineering writeup says cch is a 5-character integrity hash over the serialized request body, with Claude Code first writing cch=00000 and Bun's native layer replacing it later captured billing header reverse-engineering article.
• GitHub also disabled forks that did not contain the leaked proprietary Claude Code source, including repos that only carried edits to Anthropic's open source SDK fork lineage fork takedown example original notice context.
• Anthropic then filed a partial retraction saying all disabled repositories should be reinstated except nirholas/claude-code and the 96 fork URLs explicitly listed in the original notice retraction link notice history.
• The weirdest detail is GitHub's own note on the original notice: because the reported network exceeded 100 repositories, GitHub processed it against an entire 8.1K-repo network, which is how the blast radius got far beyond the listed targets 8.1K network note history thread.

You can read the original DMCA notice, the partial retraction, the reverse-engineering writeup on Claude Code's request signing, and the merged free-code pull request. The reverse-engineering post is especially good because it traces the hash from JavaScript placeholder to Bun-native patching. The GitHub DMCA pages are worth reading side by side because the retraction is narrower and much more precise than the takedown execution that developers saw.

The new technical detail in the fallout was not the leak itself, it was the billing header Anthropic had added to Claude Code requests. Paolo Azzan says the scheme is now fully reversed, and his screenshot shows the header format in the wild: x-anthropic-billing-header: cc_version=...; cc_entrypoint=cli; cch=...; captured billing header.

According to the reverse-engineering writeup, Claude Code's JavaScript builds that header with a dummy cch=00000, then Bun's native runtime swaps in the real value during request processing. The article says the server rejects bad values with the same error users had already seen around fast mode.

The merged PR turns that analysis into code. Its summary says cch is an xxHash64-based integrity hash over the serialized request body, masked down to 20 bits and rendered as a 5-character lowercase hex string in the billing header.

The implementation details line up across the PR and the writeup:

• Build the request body with the placeholder cch=00000
• Hash that serialized body with xxHash64 and the extracted seed
• Keep the low 20 bits
• Left-pad to 5 hex characters
• Replace the placeholder before sending the request

That matters because the reverse-engineered version no longer depends on Anthropic's shipped CLI binary. The PR explicitly frames it as enabling third-party clients to interoperate with the subscription-backed request path that Claude Code had been gating.

The legal cleanup broke into the open when maintainers of public forks started getting GitHub takedown emails for repositories that, by their account, did not include the leaked proprietary source. Theo said his disabled repo only contained a PR that edited a skill from weeks earlier fork takedown example.

That matched the confusion in the public paperwork. Gergely Orosz pointed out that Anthropic's original notice did not mention the open source Claude Code SDK, yet GitHub still disabled forks tied to that network notice context.

By the evening, the paper trail changed. The April 1 retraction says Anthropic was retracting the March 31 notice for every disabled repository except https://github.com/nirholas/claude-code and the 96 fork URLs individually listed in the original filing.

That wording matters because it narrows the target set to the specific repo and explicit forks Anthropic says held proprietary code. It also supports the claim in the tweets that the broader fork takedowns were not the intended end state, even if developers still had to wait for GitHub to unwind them.

The most concrete explanation for the blowup lives inside GitHub's copy of the original notice. GitHub added a note saying that because the reported network was larger than 100 repositories, and the submitter alleged that most forks were infringing to the same extent as the parent, it processed the takedown against the entire network of 8.1K repositories, parent included.

That note introduces a second gap in the story, separate from the cch reverse engineering. One gap was between Anthropic's proprietary client and third-party clients. The other was between the URLs Anthropic listed and the much larger repository network GitHub says it actually processed.