Skip to content

Castwright v1.12.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 10 Jul 14:02
v1.12.0
c76e7b8

A more honest transcript, twice over. A new deterministic evidence pass now catches and corrects tag-proven speaker mistakes before you ever see them, replacing the analyzer's own inflated self-reported confidence with an honest, derived one — and Script Review's findings, like your cast's own state, now survive a reload instead of evaporating the moment you look away. A couple of reliability fixes round out the polish: a stuck Qwen voice design now recovers on its own, and the Pinokio one-click installer — broken by an incomplete first fix — actually installs now.

✨ Headline features

🗣️ Deterministic dialogue-structure attribution (new) — srv-59

A structure-evidence pass now corrects tag-proven speaker mistakes before you ever see them, and replaces the model's own inflated self-reported confidence with an honest, derived value.

  • A new pure-code server/src/analyzer/dialogue-structure/ module family (per-language convention tables for en/es/fr/de/ru) slots into attributeChapterStage2: a structure parser derives tag-name/tag-pronoun/alternation/unanchored speaker evidence per paragraph, an aligner matches it against the stage-2 model output (tolerant of glyph/whitespace drift), and a cross-examiner replays every sentence against that evidence with confidence values drawn from one CONFIDENCE constants block instead of the model's own — measured on a 14,065-sentence Russian book, every sentence self-reported confidence ≥ 0.8, so the existing < 0.75 low-confidence triage view never fired despite heaps of real ambiguity.
  • Two hard invariants: a tag-name attribution is never overridden by anything, including the new default-on 'local' escalation pass over unresolved conversation windows (analyzer.structure.escalation); a dash-continuation sentence inside a speech span is exempt from the narrator-default demotion, closing a bug in the shipped plan-221 Wave A heuristic this engine absorbs.
  • Engine-off (analyzer.structure.enabled, default on), an unsupported book language, or a chapter's alignment rate falling below its floor all degrade to byte-identical pre-engine output. Two new additive state.json fields (analysisProvenance: engine/model/timestamp; a per-run structure-engine report: aligned%/confirmed/corrected/flagged/escalated) close a forensics gap — nothing previously recorded which analyzer produced a given analysis.
  • New Castwright-owned dash-dialogue Russian fixture plus full unit/integration coverage across parser/aligner/cross-examiner/escalation/ windows. On-box acceptance against the full Ночной дозор book is owed post-merge. See plan 247 (#1482, srv-59, Closes #1471).

💾 Script Review results now persist (new) — fs-58

Findings from a script review used to live only in the moment — reload the page, close the tab, or click away, and they were gone, whether you'd acted on them or not. They now survive across a reload, and closing the results modal can never silently discard them.

  • The server checkpoints each chapter's findings into a per-book on-disk ledger as a review runs — a mid-review reload reattaches to the same server-side job instead of losing it.
  • The modal's close paths now dispatch a non-destructive hide; only an explicit, confirmation-gated "Dismiss all" discards findings. A badge + re-run confirm gate also stop a fresh review from starting blind over a chapter with unresolved findings already sitting untouched.
  • Fixed a data-loss bug found while shipping this: a single-chapter (or partial) review run could silently wipe OTHER chapters' still-unresolved findings out of the in-memory bucket, and mount-time hydration could force-reopen a modal the user had explicitly hidden. setReview now merges by chapter instead of a whole-bucket replace; hydrateBucket preserves an existing bucket's visible flag. See docs/features/INDEX.md's fs-58 entry for the full design (#1479).

🎬 Caption/SRT export (new) — fs-52

Any finished book can now export .srt/.vtt captions — line, sentence, or word granularity, whole-book or per-chapter — as a new captions BookExportJob format (fs-52, #PR).

  • Line and sentence captions are pure metadata reconstructions over segments.json + manuscript-edits.json (server/src/export/build-captions.ts, caption-cues.ts, manuscript-sentences.ts) — no new render-time data, no synthesis-path change. Line cues cap at 7s / 200 chars and always close on a speaker change; sentence text is always read from the live manuscript edits, never the (potentially stale) analysis cache.
  • Word captions add a new sidecar capability: /transcribe accepts an X-Word-Timestamps header that requests faster-whisper's per-word alignment and switches on condition_on_previous_text for cross-window coherence — a distinct decode profile from the existing ASR content-QA caller, which never sets it, so that path is untouched. Always one whole-chapter Whisper call, run at export time only (never during synthesis).
  • De-dupe and filenames are now caption-variant-aware: the export queue's same-format revocation and filename derivation key on captionFileFormat/captionGranularity/captionScope, not just format, so the 12 caption variants (2 file formats × 3 granularities × 2 scopes) no longer collide.
  • A staleness guard protects against captions drifting from the audio. A segment whose textHash no longer matches its current manuscript text fails the export outright with a "regenerate this chapter" error; a segment rendered before textHash existed (pre-#1105) instead sets a non-fatal, persisted job warning surfaced as an amber caption on the export-queue row — captions still ship, but flagged as unverifiable rather than silently trusted.
  • New Captions tile on the Listen view's "Or download a file" section (src/modals/export-audiobook.tsx, src/components/listen/listen-download-section.tsx). See plan 2026-07-10-fs52-caption-srt-export.md (fs-52, Closes #975).

🔊 Generation quality & reliability

  • A Qwen voice design could intermittently fail with a blank "Internal error." _load_qwen_model (server/tts-sidecar/main.py) has hit a recurring, intermittent NotImplementedError: Cannot copy out of meta tensor when moving a freshly loaded VoiceDesign model onto the GPU since 2026-05-26 (86 occurrences) — some submodule the checkpoint doesn't cover lands on the meta device, and every reload attempt in the same sidecar process keeps failing identically; only a fresh process has been observed to load cleanly. The load failure now schedules a sidecar self-recycle on a dedicated exit code (_MODEL_LOAD_FAULT_EXIT_CODE = 44, deliberately distinct from the memory-watchdog's restart code so it can't trip the Node supervisor's unrelated "device too small" streak-trip) — the failing request still surfaces its real error, but the next design attempt hits a clean process instead of failing forever. This is a mitigation, not a root-cause fix; the underlying transformers/qwen_tts interaction is tracked in #1469 (#1470).

🏗️ Under the hood

  • Script Review's in-flight jobs can now be cancelled, and a reload can no longer silently start a duplicate re-review. Two narrow, deliberately-accepted gaps in the persistence work above (#1479) are closed. A new book-level POST .../script-review/cancel reuses each running job's existing AbortController; a mid-chapter cancel now skips the checkpoint for the in-flight chapter entirely rather than persisting a partial result the ledger couldn't distinguish from a fully-reviewed one, and the SSE stream now ends with a terminal cancelled event instead of hanging with no terminal event at all. A new join-only POST .../script-review/attach closes a TOCTOU race in the reattach-on-reload path: if the job finished in the gap between the client's GET /state and its join, the route now 404s instead of falling through to create a fresh job, and the client falls back to a plain ledger re-read instead of silently re-running the whole review. attachToRunningReview deliberately still never dispatches the shared clear action itself, preserving the persistence work's own concurrent-reattach fix (#1481).
  • The off-roster reattribute confirm form could leak a field across a multi-op batch. CreateCharacterForm's name/gender/ageRange fields are seeded via useState(initial ...), which only runs on first mount; ScriptReviewDiff's confirm-queue overlay (src/components/script-review-diff.tsx) rendered the same JSX element across every confirm.index advance with no key, so React reused the instance and a field typed for op N could survive into op N+1's own form — and, if unedited, reach op N+1's create/reassign call. The form is now keyed on the op's own identity (opKey) so it remounts, and its state resets, on every confirm-queue advance (#1480).
  • The Pinokio one-click installer's menu was fundamentally broken, across two attempts to fix it. First surfaced during on-box acceptance (#822): root pinokio.js used CommonJS syntax (require/__dirname/module.exports) under the root package.json's "type": "module", throwing a ReferenceError (bug #1458) — "fixed" by rewriting it as genuine ESM (import/export default), on the assumption that Pinokio's runtime dynamically import()s app scripts. That assumption was itself wrong (bug #1484): Pinokio's kernel (pinokiocomputer/pinokiod, kernel/loader.js, requireJS()) actually loads scripts with a plain synchronous require(filepath), which cannot load an ES module at all — so the install screen still never loaded, for a different reason. Root pinokio.js is now genuine CommonJS again, matching every other pinokio/** script, and the root package.json is "type": "commonjs" (was "module") to make that load correctly; eslint.config.js — the only other bare .js file at the repo root relying on ESM syntax — is renamed to eslint.config.mjs to keep working under the new default. The regression test now require()s the entry via createRequire instead of dynamically import()ing it, so it actually reproduces Pinokio's real loading path (#1458, #1484, #1485).
  • FlashAttention-2 detection is no longer Windows-only, and the stale Windows pin is no longer the only path. install-qwen3.mjs --flash-attn now checks whether flash_attn is already importable in the sidecar venv on ANY platform/accelerator profile first — if so, it just reports how to activate it (QWEN_ATTN_IMPL=flash_attention_2) rather than attempting a reinstall. On Linux, when it isn't already present and the standalone NVIDIA CUDA Toolkit (nvcc) is on PATH, the installer now attempts an unpinned pip install flash-attn --no-build-isolation build; without the toolkit it skips cleanly instead of burning a doomed compile. The Windows pinned-wheel path (still cp311-only; the cp312/torch2.11/cu128 real-stack fix is tracked on side-22, #1001) is unchanged. (side-21, #1000)
  • The library welcome subhead no longer hardcodes "book seven." LibraryChrome (src/components/library/library-chrome.tsx) claimed a character "carries through to book seven" regardless of how many books are actually in the library — now reads "book after book," which holds true at any count (#1493, Closes #1461).

Full changelog: v1.11.0...v1.12.0

Assets 8
Loading