Review of talk-mnemo-bot PR #39: attachment hardening (PDF bounds, pixel limit, name escaping, refusal admission, no fileĀ retention)

From mnema-talk v3.3.13, rendered on this site.

Two rounds on mnema-xyz/talk-mnemo-bot. The first reviewed the five changes as first pushed; its should-fix findings and one nit were fixed in the same pull request and checked in the second. Released in Mnemo 0.1.14.

What the pull request fixes (found in the 2026-11-19 audit of the attachment feature, #22, #29, #37):

  • PDF text extraction used lazy regular expressions whose cost grows with the square of a hostile input (a 560 KB file of repeated stream openers held the event loop for about 7 s), inflated streams without an output limit (255 KB of deflate became 256 MB), and never cut the extracted text. One member's PDF could stop or restart the bot with a provider other than Anthropic.
  • The text of a consented attachment was kept in the thread history and the answer record for 24 hours and sent again with every later question in the thread and with Regenerate, without a new consent.
  • Refusals of an attachment were posted before the rate limit.
  • Images were decoded up to sharp's default of about 268 megapixels.
  • Member names went into the Markdown consent prompt unescaped.

Round 1

Verdict: MERGE AFTER FIXES. All five changes improve on main and each one's tests fail when it is reverted. Three should-fix items.

Run: npm test 159 pass, npm run lint and npm run typecheck clean (the lint config enforces the 400/800/80/15 limits). Each change reverted on its own in a scratch copy: its tests fail (old extractor: 40 s for unclosed stream openers, 1.7 s and 540,000 characters for a 64 MiB stream).

PDF probes up to 10 MB: nothing super-linear; the worst case about 1.5 s per admitted question (2,527 streams each just over 4 MiB: 1 s; eight 4 MiB streams of nested (: up to 1 s and 417 MB heap before collection).

Name escaping was run through copies of the Talk parser (blocks.ts, inline.ts, rules.ts, mentions.ts): links, autolinks, headings, quotes, code and spoilers stay text; a code break-out is not possible.

Refusal admission: the mention is admitted once, Yes once; admit holds nothing back. Retention: thread memory and AnswerRecord.question hold the bare question; an answer about a file has no Regenerate button and a forged one is refused before admission; the consent record holds the message, not the file, so every use needs a new consent.

Findings:

  • S1 (should-fix) src/bot/consent.ts:32: an escaped @all, @alice or <:name:id> still renders as a mention pill or custom emoji, because the renderer's text rules run after escapes resolve. Visual only (no pings).
  • S2 (should-fix) src/bot/file-text.ts:153-160: strings built one character at a time, about 32 bytes of heap per input byte; out of memory at a 96 MB heap.
  • S3 (should-fix) src/bot/file-text.ts:29-30: a stream skipped for its size was not charged against the 32 MB total.
  • Nits: /details did not say a file was sent; tool inputs taken from a file stay in admin-only /details for 24 h; a message lookup runs before admission; the name cut could split an emoji; an unclosed ( ends the scan of its stream; the README named only the 4 MB limit.

Not checked: real-world PDFs beyond the fixtures, Talk's server-side mention parsing, other renderers, provider-side prompt caching, the integration tests.

Round 2 (commits decea7d, 332a3d6, f52b388)

Verdict: MERGE.

Run on the final head: npm test 169 pass, lint and typecheck clean. Each round-2 change reverted to the round-1 head in a scratch copy: its tests fail (7, 2 and 2 tests).

  • S1: a word joiner follows each escaped @, < and [ (also wiki links); the parser copies no longer match the mention, role, custom emoji or wiki link rules. Names are cut by whole graphemes; ZWJ stays.
  • S2: strings are built from slices with a per-string cap; TJ arrays are capped. Nested (: 1,019 ms and +478 MB before, 90 ms and +21 MB after; the combined worst case 5 to 8 ms with 8 MB heap, and it runs at a 64 MB heap.
  • S3: a skipped stream counts as 4 MB; 2,527 over-cap streams take 9 ms.
  • /details adds "the attached file" only for answers about a file.

Remaining nits: a PDF with eight or more large images before its text loses the text after them (the README describes the trade-off); the round-1 nits on /details tool inputs, the lookup before admission and an unclosed ( are unchanged.

Something wrong on this page? Tell us.