[owengo]

jsonStringLength walks every character against MAX_STRUCTURED_CHARACTER_WORK (1M), and compactToolContent measures the same content more than once with one shared allowance (originalChars, then the normalized length). A data URL over ~500k characters exhausts it. The content then measures as Number.MAX_SAFE_INTEGER, and the atomic block is omitted whatever the cap allows.

• A base64 data URL needs no JSON escaping, so its JSON length is length + 2. jsonStringLength now returns that for a string that fully matches a base64 data URL and is no longer than HARD_MAX_TOTAL_TOOL_OUTPUT_SIZE, the ceiling isProviderImageUrl already applies to inline images.
• Every other string, including a data URL with an escaped character or one above that ceiling, keeps the bounded walk. The caller's cap still applies: an image larger than the cap is still omitted.

projectArtifactPayload / projectAnthropicArtifactContent
  └ appendBoundedContentSegment
      └ compactToolContent              one StructuredWorkState for every measurement
          ├ estimateStructuredChars     originalChars
          ├ estimateStructuredChars     normalized length
          └ estimateStructuredChars     each atomic block against the atomic budget
              └ jsonStringLength        base64 data URLs no longer spend the allowance

Related to #602 (the same report) and #603 (the same approach). Their author withdrew both without review, and the behavior is unchanged in 4.2.1. We have run the equivalent change as a patch on 4.0.4 in a LibreChat deployment since 2026-11-04.

Verification

Check	Result
New regressions on main at 708c9118: 600k and 3.2M-character images kept under an 8M cap, and both artifact projections keep a 600k-character image	3 failed before the fix, passed after
Guard regression: a data URL with a ", a data URL above the ceiling and a plain 1.1M-character string still exhaust the walk	Passed before and after
npx jest src/utils src/messages src/graphs/__tests__/Graph.contextOverflow.test.ts	1,338 passed, 1 skipped across 45 suites
Full unit suite with CI's exclusions (--testPathIgnorePatterns as in validate.yml)	8,007 passed, 68 skipped; 328 suites passed, 21 skipped
npm run typecheck, ESLint and sort-imports --check on the touched files, git diff --check	Passed
npm run build:dev, npx tsc -p tsconfig.build.json, npm run check:circular-deps, test:circular-deps, test:light-entries, test:typecheck	Passed

Each measurement of a data URL now costs one native regex pass instead of a character walk capped at 1M: about 14 ms for a 3.2M-character URL on our machine. The ceiling bounds that cost. The changed lines follow the pre-commit Prettier and ESLint pass. Existing formatting drift elsewhere in formatAgentMessages.test.ts is left untouched.

Not run: live provider requests or Langfuse export. The change only alters how a string is measured, not what is sent or traced.