Skip to content

v6.4.4

v6.4.4 — Honor cron allowedTools end-to-end so fb-* crons actually post
@Djtony707 Djtony707 tagged this 23 May 02:54 
Live observation during the v6.4.1 audit caught all 5 Facebook crons firing
on schedule and producing zero posts. The model was being asked "Call fb_post
ONCE" but the agent loop sent tools=[] to Ollama, so the model hallucinated
the tool-call syntax inside its text body. HallucinationGuard correctly
rejected it (good), but no real post went out.

Three stacked bugs:

1. executePrompt() silently dropped allowedTools. The `as Parameters<...>`
   cast let it pass through, but processMessage's overrides type had no
   allowedTools member. The comment said it all: "allowedTools is plumbed
   but not yet honored end-to-end — logged here so a future hardening pass
   can enforce it." This is that pass.

2. allowedTools is stored as a comma-separated STRING ("fb_post,fb_read_feed")
   for legacy cron entries. The truthy check passed (strings have .length)
   but downstream filters expected an array. Normalize both shapes.

3. fb-afternoon misclassified as 'content' instead of 'social' because the
   content rule's `\bwrite.*comparison\b` matched "Write an afternoon
   COMPARISON" first. Swap social/content rule order — social wins for any
   prompt mentioning Facebook explicitly.

Fix: cron.ts normalizes string→array. agent.ts adds allowedTools to
processMessage overrides; when present, short-circuits the persona /
brain / toolSearch / dangerous filter chain entirely and binds EXACTLY the
requested tools (with [AllowedTools] override bound N/M log + warn on missing
tool names). pipeline.ts swaps rule order.

Hoisted allToolsBackup + toolSearchEnabled above the new short-circuit so
both branches can reach them (used downstream by runAgentLoop).

7848/7850 tests pass. Typecheck clean. UI builds clean.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Assets 2
Loading