ttlequals0/MinusPod v2.100.23

2.100.23

on GitHub

Added

• utils/segments.py helpers for validating, normalizing and joining word-timed segments.
• Stats: cache read and write tokens on model usage, per-run totals and contributing calls. A cache hit rate (reads over input) shows on the model usage table, the contributing calls panel, the per-run phase breakdown and the LLM token tile (#897).
• Pricing: model_pricing stores a cache read rate from the OpenRouter API or LiteLLM, filled in from a later source when the primary one lacks it. Operator overrides accept cacheReadCostPerMtok, and the Settings pricing form has an optional cached input field. Estimated costs on OpenAI-compatible and OpenRouter rows bill cached input at that rate (#897).

Changed

• README rewritten as a short overview: one-line feature bullets that link to the topic docs, a quick start of three commands, and the setup, password and release-channel detail left to the installation and security docs.

• Docs deduplicated. Each topic is explained once in its own doc and other docs link to it instead of restating it, and upgrade history narration is rewritten as current behavior. Stale glossary entries, the podcast search default, old button labels and a dead anchor are fixed.

• openapi.yaml carries a fixed placeholder version. The served spec already substitutes the running version from version.py, so the file no longer needs a bump on every release.

Fixed

• Ollama: a call whose prompt Ollama silently truncated to fit num_ctx now fails instead of returning an answer written without the instructions. The native client compares prompt_eval_count with num_ctx/2 and the prompt's own size estimate. The error names num_ctx, a lost detection window is counted under the new context_overflow loss class, and a chapter run degraded by it says so in its degradation reason. (#893)
• Docs: the Ollama context window advice now recommends 32768 or higher. Chapter prompts with the default 2700 s windows run 18k to 23k tokens, so the old "8192 or higher" advice left every chapter call truncated. (#893)
• Transcript text built from word timings no longer renders split Whisper tokens as "AT &T .com"; words are normalized (merged, space-prefixed, sorted) as soon as a backend returns them (#899).
• OpenAPI: the per-episode cost runs endpoint and the processing-admission pause endpoint are documented, and descriptions drop dash asides, version history and two stale README pointers.