Script Filter Plugin

The Script Filter plugin removes runs of non-Latin characters (CJK, kana, hangul, …) that a model sometimes leaks into otherwise-Latin output. It strips them live during streaming (replacing each run with a placeholder) and, when the turn finishes, optionally translates the removed fragments back into the surrounding language and revises the message in place.

Configuration

plugins:
  - module: codumentor.plugins.script_filter
    class: ScriptFilterPlugin
    priority: 70
    args:
      enabled: true
      scope: both              # both | content | thinking
      placeholder: "□"
      translate: true
      target_language: auto    # 'auto' = match the surrounding text
      request_timeout_seconds: 20.0
      defer_translation: true

Parameters

• enabled (optional): Strip characters at all (default: true)
• scope (optional): Where to strip — "both", "content" (visible answer), or "thinking" (reasoning) (default: "both")
• placeholder (optional): Shown in place of each removed run during streaming; "" deletes with no marker and disables translation (default: "□")
• translate (optional): Translate removed fragments when the turn finishes (default: true)
• target_language (optional): Language to translate into; "auto" detects the dominant surrounding language (default: "auto")
• request_timeout_seconds (optional): Hard timeout for synchronous translation (default: 20.0)
• defer_translation (optional): Run finish-time translation as a background job (with retry) instead of blocking the turn; falls back to synchronous translation if no job queue is available (default: true)
• Advanced: model_role, temperature, max_tokens, coalesce_adjacent, translation_max_attempts, translation_retry_backoff_seconds, and custom unicode_ranges / pattern for the character set to strip.

How It Works

The plugin has no agent tools — it is pure content transformation via two hooks:

1. onLLMStream (priority 20, early): replaces target-character runs in each streamed delta with the placeholder and caches the removed text, keyed by the turn's run_id.
2. onBeforeSendResponse (priority 70): when the turn finishes, translates the cached fragments and edits them back into the final message (per segment for multi-segment turns).

When defer_translation is on, the finish-time translation runs as a script_filter_translation background job: it retries with backoff and rewrites both the displayed message (an SSE revision) and the LLM-history placeholder rows.

Notes

• The placeholder briefly flashes during streaming; the translation lands after the turn completes, as a post-hoc revision.
• If translation fails (or placeholder is ""), the placeholders simply remain — the turn is never blocked or corrupted.
• Subagent output is left untouched.
• A settings widget toggles stripping/translation and sets the placeholder and target language.
• Finish-time message revision relies on the same machinery as the Permalink plugin's response correction.