Skip to content

v0.14.4

Choose a tag to compare

View all tags
@github-actions github-actions released this 18 May 16:01
· 11 commits to main since this release
v0.14.4
c927686
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Highlights

  • Preemptive chunking at every high-severity LLM call site. Large docs sites used to occasionally trip ErrTokenBudgetExceeded partway through analysis. The reactive fitContentToBudget / chunkObservationsToFit helpers are gone; in their place a new internal/chunker package splits content on heading → paragraph → sentence boundaries while protecting code fences, tables, and lists, applied before the call ever leaves the box. Coverage: internal/chunker 93.9%, internal/analyzer 92.3%.
  • Call sites updated: analyze_page, screenshot_gaps, the drift investigator + judge, and synthesize.
  • Drift investigator compression. The system prompt is rewritten to be shorter, and the full symbol / page lists are now exposed through two new tools the model can call on demand instead of being inlined into the prompt.
  • synthesize map-reduce fallback. Adds per-page summary compression with a map-reduce fallback path for very large docs sites that would otherwise exceed the synthesis budget.
  • Cleaner overflow logs. Any remaining ErrTokenBudgetExceeded falls through to a single WARN-level log with full site context, instead of bubbling up as an unrecognized error.

See .plans/CONTEXT_OVERFLOW_AUDIT.md and .plans/CONTEXT_OVERFLOW_REMEDIATION.md for the call-site audit and full design.

What's Changed

  • fix(analyzer): preemptive chunking at HIGH-severity LLM call sites by @britt in #103

Full Changelog: v0.14.3...v0.14.4

Contributors

britt
Assets 7
Loading