Strip whitespace + ensure URL starts with http(s).

Persist overrides to disk (atomic write via temp file + rename).

* Run the full panel over every provided chapter. * The caller supplies chapters (id/number/title/text) and AI callables.

* Score a project for priority-based selection. * Higher score = picked first.

* Search Google Books API (free, no key required for basic search). * Returns books related to the query.

* Search Wikipedia using the MediaWiki API. * Free, no API key, no CAPTCHA, returns titles + snippets.

* Self-improvement: After a project completes, analyze the outputs * and store actionable insights for future writing.

* Wire up autonomous capabilities. Called after project engine and AI are ready.

* Replace the domain allowlist and persist to disk.

Wire a callback that returns the names of skills already in the catalog * (so we don't generate duplicates).

* Inject the gateway's message pipeline so the engine can execute project * steps through the full AI stack. Called from index.ts after the gateway

* Inject the subset of gateway services the step-execution hooks use * (quality judge, activity log, heartbeat, TTS, personas, aiRouter).

* Skip a step

* Split markdown content into chapters based on # or ## headings.

Convert SRT to plain text — strip timestamps and indices.

* Returns true if the step requires the FULL manuscript in context (not a * truncated excerpt). Revision-apply rewrites AND consistency checks must

* True if this step should pull ARCHIVAL excerpts: revision / consistency * steps (by taskType or phase) or ANY step of a book-production project

Remove any HTML tag that KDP doesn't accept.

Strip all HTML tags for the plain-text fallback.

Whisper API transcription. Downloads audio-only, transcribes, deletes audio.

* Translate a single chunk with one retry. Never throws — on double failure * it returns { failed: true } so the caller can emit an error marker an

Update allowed users on a live bridge (called when dashboard saves users)

* Update autonomous configuration at runtime

* Lightweight frontmatter validation. Warns (once per skill) about * missing/malformed fields; never throws — the skill still loads with * whate

* Wrap a title into lines of roughly `maxCharsPerLine` without breaking words. * Aims for 1-3 lines. Single-word titles return as one line.

* Analyze a sanitized review set and produce the intelligence report. * All clustering is keyword-based (no AI embeddings) so it's deterministic

* Run the full craft analysis on a set of chapters.

* Audit a project's promises against current progress. Returns a report * highlighting at-risk promises (still open near the end of the manuscript)

* Auto-advance word-count goals using current project word counts. * Call this after step completions / via heartbeat.

* Fully autonomous mode: loop over ALL active steps of a project, executing * each through the message pipeline. Long-running — can run for many mi

Async helper: poll every `intervalMs` until a terminal state or timeout.

* Broadcast a message to all Telegram users. * Used by routes for conductor status updates.

* Build the full static site to workspace/website/<slug>/. * Returns the output directory path plus a list of pages and warnings.

* Format the user model for system-prompt injection. Keeps the size small * (~300 tokens) so it doesn't dominate context.

* Build a context block for injecting into AI prompts when this persona is assigned to a project.

* Build the user message for project step execution. * Injects uploaded manuscript DIRECTLY into the user message so the AI can't * miss it. For

* Clean up old audio files (older than 24 hours).

* Clean up old images (older than 7 days)

Compute progress for all active goals (dashboard convenience).