Skip to content

v0.13.0: path input, errors and warnings only (engine 0.16.7)

Choose a tag to compare

@Ouranos27 Ouranos27 released this 23 Oct 15:55
· 11 commits to main since this release

Less of the email, and less about it, crosses the model's context. Measured
with Claude Opus 5.5, GPT 6 Luna and Gemini 3.8 writing five emails each,
every tool call logged: all three passed a path without being told to.

Added

  • path instead of html on every tool that takes a source (and
    beforePath / afterPath on diff_emails). The server runs beside the
    files an agent edits, so the email no longer travels through the model's
    context on each call of a lint, fix, re-check loop. On a path, fix_email
    points at the file instead of repeating the source in its prompt. Only email
    sources are read (.html, .htm, .mjml, .jsx, .tsx, .vue, checked
    after symlinks, size-capped before reading); the extension sets format.

Changed

  • Compact JSON responses. Pretty-printing was about 30% of every response.
    Together with path, an analyze, fix, audit, re-analyze loop on the
    Cerberus fixture went from 396K to 278K characters (-30%).
  • The summary lists errors and warnings, and counts info. analyze_email
    and audit_email return infoFindings (a count) instead of listing safe
    fallbacks; on a typical agent-written email that was 35 of 37 entries. Each
    audit reply came back 46% smaller for GPT 6 Luna and 60% for Gemini 3.8.
    detail: "full" still returns everything, info included.
  • Engine 0.16.7 (was 0.16.4). A body background repeated with
    !important, or by the rule that also paints the wrapper, is no longer
    reported lost; Gemini 3.8 had audited each email about six times on it
    (0.16.6). The Outlook Classic preview matches real Outlook 2016 and 2019
    screenshots: rem padding dropped, mso-font-alt honoured, * ignored
    (0.16.7). Scores do not change.