self.md radar — 2026-10-26
129 MB for one file-write call
check the client before upgrading to langchain-core 1.7.0. the tool-argument stream now sends slices instead of accumulated snapshots.
1. langchain-core 1.7 stops resending the arguments so far
one 64k-character write_file call, streamed in 16-character chunks, produced about 129 MB of events under the old bridge. the merged change puts the new figure at 0.5 MB. those are event payloads for the stated example, not model tokens.
version 1.7.0, released October 22, emits tool-call-delta slices instead of block-delta snapshots. clients consuming the raw stream must follow the new event type. .tool_calls.get() and .output are unchanged; consumers that still need the old snapshots can pin langchain-core<1.7.
the PR says to upgrade JavaScript and Python clients first, then upgrade langchain and langchain-core together on the server.
reading: release and compatibility notes
2. PaperFormalizer asks before weakening a main result
five of seven results in one paper needed revision; in another, 15 of 29 did. those are the authors’ reports from using PaperFormalizer on three of their own papers, not an estimate of how often published mathematics is wrong.
the pipeline starts from LaTeX, reviews the statements and builds Lean proofs. a repair that substantively changes a main result waits for the author. smaller repairs can be applied automatically, with the original and revision exposed for review. the author can reject them or report that the agent misread the paper.
in the seven-result case, the authors say five did not hold as printed. they accepted weaker statements, and all seven revised results were proved in Lean. the run also recorded 94 clarifications.
reading: PaperFormalizer preprint · code and setup requirements
3. Six unknown outcomes pass a recovery test
a refund commits, its acknowledgement disappears, and the person who requested it loses write permission. sending it again and checking what already happened need different permissions.
Permission to Retry tests recovery policies across 336 scripted histories on native LangGraph and Temporal hosts. per host, retained-ledger reconciliation passed all 42 histories: 21 checked completions, 15 denials and six unknown outcomes whose reservations stayed held. the test does not call those six completed.
cached retry, revalidation before retry and stable-key retry passed 30, 36 and 31 histories respectively. these are application policies written for the study, not framework defaults. the paired host outcomes were identical. decisions were scripted rather than generated by a model.
reading: study, provider contracts and exclusions
4. A crop through the middle
Simon Willison wanted to remove an irrelevant prompt from a long image stitched together with Tailor on his iPhone. keeping the top and bottom ruled out an ordinary edge crop.
his October 25 tool, Seam Cutter, lets you place a horizontal or vertical seam, slide the sides relative to each other and export a PNG. the source uses a canvas and browser download for export.
reading: why he built it
keep the unedited screenshot too. removing a prompt changes what the next reader can infer about the answer.