Build {id: node} from the tree's node list. Silently skips entries missing an id so partial trees from mid-edit clients still work.

Best-effort: a token is sentence-initial if it's the first token or the previous token (stripped of casing) ends a sentence. Since we dropped

Treat a token as a possible proper noun if it starts with an upper Unicode letter and contains at least one other letter. We only use this as

Read the Hermes user config without pulling in PyYAML. The file is small and we only need a couple of strings; a hand-rolled line-walk avoids

Return (franchise_slug, matched_tokens) — lowest-token winner is fine because the heuristic is "any hit means known franchise".

Walk parentId upward. Excludes the node itself. Root first, target last's parent.

Walk canonChildId from root until the spine terminates.

BFS from node_id. Excludes node_id itself. Respects depth_limit when set.

All node ids that share sceneId with node_id. Order follows input order.

All children of the node's parent, excluding the node itself.

Return a diagnostic report. Never raises.

Synchronous HTTP POST to the FastAPI render service. Uses urllib so the tool stays dependency-free — the hermes-agent venv doesn't need reque

Best-effort duration probe via ffprobe. Returns None if ffprobe is unavailable so consumers without ffmpeg don't crash on this.

Where Noustiny-style /cache/X paths might be served from. Set NOUSTINY_PUBLIC_DIR to your Next-style /public directory. Generic enough for an

Read HF token from the canonical ~/.cache/huggingface/token path where `huggingface-cli login` writes it. Stdlib only.

Reconcile an LLM-emitted name string with the canonical key under which a character was registered. See `resolveCharacterAlias` in web/lib/ch

Return {base_url, api_key, model} for the LLM call. Resolution order: 1. Explicit env vars HERMES_DIRECTOR_BASE_URL / HERMES_DIRECTOR_API_K

Prefer explicit rootId, fall back to the single parentless node.

Scan one beat's body; update running registries. We scan **body only** because titles are noisy: they capitalise non-proper-noun words (fonts

Run diarization on ``audio_path``, pick the dominant speaker's longest contiguous monologue, and return the (start, end, confidence) tuple to

Compact statistical fingerprint of the canon. The skill uses this to infer *register* (clipped action vs long lyrical prose), *cadence* (what

Return an absolute HTTP URL for a ``/cache/refs/...`` or * ``/cache/images/...`` path so Nous Portal image models can fetch it. * Disk paths live

Build N visual prompts for N new sibling beats in a single Haiku call, * then fire N image-gens in parallel. Saves ~60% of the per-Enter * Claude

Build the multi-turn chat history that makes `expandNode(nodeId)` read * as a real back-and-forth conversation instead of a stateless single- * sh

Return all cached voice records, newest-last by last_used. Used by the cleanup tool's LRU evictor.

Bump last_used on cache hit so the LRU evictor preserves the most-recently-active voices.

Walk the canon from root up to (but excluding) ``nodeId`` and return * each ancestor's {id, title, body, mood}. Feeds the narrative * context bui

Runs the character-sheet-builder skill once per story. Returns a * JSON array `[{name, description, portrait_prompt}, ...]` the client * uses to

Runs the story-copyright-detector skill (Claude reasoning) but falls * back to the deterministic TS mirror if the gateway is slow / down. * Envelo

Best-effort cleanup of a stored voice. Returns True on success. Errors are logged but never raised — a leftover slot is annoying but not a re

* Walk a chat completion response looking for the noustiny_storybook * tool call result. Hermes's agent loop runs the tool itself, but the * SHAPE

Mood-based fallback for the ambience layer while the skill hasn't * shipped its body-aware ambience field yet. Conservative: many * moods default

* Grid of saved-story cards. Mounts in two places: * - Landing page, inline section (seed picker alternative) * - Canvas MoreMenu → "Open saved"

Batch variant — sends N beats to visual-prompt-builder in a single skill * call, expects a JSON array back in input order. Halves-plus the Claude

Pick a narrator persona + YouTube search query for a story. Args: title: Story title (or beat title) — short, ≤120 chars typical. see

Parse a natural-prose brainstorm reply. The v6 skill doesn't enforce * a format — Gemini can return any of these shapes (all valid): * * 1. LA

Walk parent pointers from `endpointId` back to the root, return * the linear path in root→endpoint order.

Deterministic map from detector policy to a storyRegister string. * Mirrors the fingerprint table in visual-prompt-builder's SKILL.md so * charact