self.md radar — 2026-11-06
an account email became a request header
the account email was there to stop Claude Code inventing author names
when Claude wrote a web request, it could reuse that address as the caller’s contact
1. Claude Code plans to remove account emails from model context
Claude Code developer Lydia Hallie says the next release will remove the account email from model context
she expects it early in the week beginning November 9; the removal is announced, not yet documented as shipped
the interim check shipped in 2.1.317
the changelog describes asking before a web or shell call sends the account address
the published privacy hook instead returns a denial when it finds the address alongside User-Agent in shell input
Hallie names a remaining hole: Claude can write a script and run it itself
an upgrade alone does not establish that those scripts cannot send the address
the reporter’s spam claim has no traced causal chain in the issue
reading: the developer’s explanation · the published shell-input check
2. Proof agents spend most of their edits on explicit gaps
in Lean, sorry lets a proof compile with a gap still in it
in a new study of nine Lean projects, definitions carrying that marker account for 59–92% of agent modifications, against 19% of human modifications
Bálint Gyevnár and Simon DeDeo replay historical repository states, ask an agent to continue, then compare its changes with the actual next human commit
they test Opus 4.6 in Claude Code, GPT-5.2-Codex in Codex and Gemini 3.1 Pro in Gemini CLI
agents also edit the edges of the proof’s dependency graph more than humans, who work more in its densely connected middle
a preselected-theorem benchmark would have made that choice of work for them
this is simulated collaboration, not people and agents working together live
the public repositories predate the tested models and may have appeared in training; the dependency graph comes from a heuristic parser
the comparison concerns which definitions get edited, not whether their proofs are correct
reading: the experiment, results and limitations
3. Kagi permits 50 searches without an account
Kagi’s November 5 release note opens the first 50 searches to visitors without registration
creating an account adds another 100
its Android home screen now shows the remaining trial searches even before someone has an account
these are trial allowances, not a promise of permanently free search
reading: Kagi’s accountless trial and Android changes
someone receiving a recommendation can try the search engine before deciding whether to register