Verify / recast now run span-natively (no text re-emission): a second differently-windowed
pass marks lines agreed/disputed and records the other answer, without changing any speaker
(~9 min vs ~49 min). Measured on the real pipeline, overruling disputed lines with a reasoning
adjudication made them worse (66.5% -> 61.4%), so that step was removed; agreed lines were 83%
right vs 48.5% for disputed ones, which is the signal used for confidence and review.
Auto-repair no longer runs after a span cast; segments carry a spanId.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- fix: base-appearance matched pair lost its concept image + metadata on reload; visual save
wrote base concept art to the record root; one save path for base and forms now
- fix: extraction no longer lets the LLM author image_prompt (source of avatar/concept drift);
avatar-only generation uses the deterministic prompt and flags an existing pair
- fix: prompt writes and Physical/Clothing edits go to the active form; review endpoint checks
each form's own pair; fingerprint parity with the browser incl. visual_* overrides
- feat: appearance forms panel (add/rename/delete/show, per-form stale flag, generate one/all)
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Split the book into narration/dialogue spans from its own quotation marks and ask
the LLM only who speaks each dialogue span (POST /api/attribute-spans), so text can
no longer be fused, dropped or duplicated and a novel takes ~10 min instead of ~4 h.
Book 1 vs the hand-corrected reference: every-word-right 87.0% -> 92.3%, type
agreement 93.6% -> 97.4%; Books 2/3 (not tuned on) 90.8% / 89.9%.
- static/js/audiobook-spans.js (pure, unit-tested) + audiobook-cast-spans.js
- core/attribution_spans.py, route, tests; scripts/score_attribution.py
- default on; localStorage ab-span-mode=0 restores the classic engine
- architecture/audiobook-attribution.md, Manual, CHANGELOG 1.32.0, TODO
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds "Edit boxes" mode to article review: drag an already-detected
region to move it, its corner handle to resize it, or press its x to
delete it -- for correcting a box the layout engine got wrong, not one
it missed entirely ("Draw it" still covers that case). Moving or
resizing opens a confirm panel; "Save" keeps the region's existing
text at its new position, and for text regions "Save & re-OCR" reads
the corrected area again and replaces the text.
New core/magazine.py update_region()/delete_region() and PATCH/DELETE
/api/magazine/{doc}/pages/{page}/regions/{block_id}. An image region
is re-cropped from the page render whenever its bbox changes, since
its saved file is exactly that crop.
Live-verified against the real ingested c't KI Wissen 2026 magazine:
add -> move -> move+re-OCR (correctly kept old text when the corrected
area had nothing readable) -> delete -> confirmed gone, repeat delete
404s.
275 tests passing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds a "Describe" action per figure in article review: sends the cropped
image to the configured vision engine (with fallback) for a type-aware
one- or two-sentence description -- a chart gets a prompt about what it
measures and its trend, a table about headings and what stands out, etc.
Editable before approval. A new "Read figure descriptions" toggle in the
reading controls speaks these where they occur; purely additive, off by
default, and a figure with no description behaves exactly as before.
New core/magazine.py describe_visual_region()/_chain() reuse the same
reliability guards as page OCR (repetition-loop detection, sentence-
boundary truncation). New routes: POST .../images/{id}/describe
(background job), PATCH .../images/{id} (manual edit), descriptions=1 on
the article-text endpoint.
Live-verified end to end against the ingested c't KI Wissen 2026
magazine. Also found and documented a real, config-dependent limitation:
the currently-configured primary vision engine is transcription-only and
returns empty content for any descriptive prompt (confirmed with three
different prompts) -- not a code defect, but description generation
needs a genuinely caption-capable engine reachable as the fallback.
251 tests passing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Connect and "Use as LLM" on the LiteLLM Gateway card now resolve that
card's own saved key (engine_api_keys[name]) server-side in
routes/docker.py, ahead of the old global llm_api_key fallback. The
browser can no longer reliably forward a saved key at all -- it's
masked to "__configured__" in every /api/settings response the moment
it's saved -- so relying on a frontend-forwarded header alone left
Connect silently using a stale global key after LiteLLM's own key was
rotated, surfacing a misleading /api/version 404 instead of the real
401.
"Use as LLM" now goes through a dedicated /use-as-llm endpoint that
saves the exact credential Connect just verified (llm_url normalized
to .../v1, llm_api_key copied from the card), rather than a
separately-typed value that could drift again.
Also documents per-user provider-key storage as an explicit item for
the deferred multi-user/identity planning pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Resumes an in-progress, uncommitted change from a previous session: the
live "page image + recognised text" preview shown during ingestion was
reworked to render formatted HTML (styled headings, a figure
placeholder) instead of a plain-text dump, but was left with a
duplication bug, dead code, inline styles, and no tests.
- Fixed duplication: a figure block's own transcribed text (usually its
on-screen UI copy) was rendered as an ordinary paragraph in addition
to the new figure placeholder box. Figure/chart/table/code blocks are
now represented only by the placeholder.
- Furniture/sidebar/pullquote/ad text is shown muted instead of being
dropped with no replacement, and is never promoted to a heading
regardless of font size.
- Factored the formatting logic out of the progress closure into a
standalone, unit-tested _render_ingest_preview_html().
- Replaced dead getattr(bbox, "y0", ...) calls (bbox is always a plain
[x, y, w, h] list) with direct indexing, moved statistics/logging
imports to module level, switched manual character replacement to
html.escape(), and bounded the preview by an accumulated line budget
instead of a blind slice that could truncate mid-tag.
- Replaced inline style="..." with CSS classes matching the panel's
existing dark console treatment.
- The frontend now always uses innerHTML (every text node is escaped
server-side), dropping the startsWith('<') content-type sniffing.
- Marked the preview panel aria-hidden -- decorative, and would
otherwise churn noisily for screen reader users on top of the
adjacent progress bar.
Verified against the real c't KI Wissen 2026 layout: a figure's raw
OCR text no longer appears twice, a real headline renders as <h3>, a
footer/pull-quote render muted, and all text is correctly escaped.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A vision OCR engine grounds a figure far less precisely than it
transcribes one. Over all 138 figure regions of a 204-page issue the
transcribed contents were right while the boxes were routinely offset,
undersized, or sitting on the neighbouring text column -- page 23's card
grid was reported at x 0.24-0.54 where the graphic runs 0.41-0.95.
snap_visual_bboxes() reduces the page render to an ink-coverage grid,
seeds a connected-component search with the engine's box, and re-joins
tiles a gutter apart so a card grid comes back as one figure. Text
blocks are excluded from the mask, which is what stops a figure
swallowing the column beside it. A candidate must be solidly inked or
framed on nearly all of its rows and columns (admitting a white-interior
screenshot), may not shrink below 30% of the seed, and reverts when two
figures land on the same graphic; anything else keeps the box it had.
Runs before the figure crop in both OCR paths, so the images shown
during playback are framed correctly too. Replaces refine_picture_
bounds(), which only recognised coloured panels and had no text fence.
Also adds "Fix figure boundaries" in review and POST
/api/magazine/{id}/figures/refine behind it, so an issue ingested before
this pass can be corrected without re-running OCR. Verified live against
c't KI Wissen 2026: 78 of 138 boxes corrected, the rest declined.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ingestion now streams each page's image and recognised text into a
console-styled preview panel (Baidu Unlimited-OCR demo style) as pages
finish, instead of a bare progress bar -- raw text is written to a
per-job ingest_preview.json file since the background-job result field
strips text-like keys for data minimization.
Split-screen review boxes and their matching list rows now show a
centred icon for the region's kind and, when a region on a shared page
belongs to a different article, a badge naming that article, so a
claimed region is never mistaken for unclaimed text.
Added a "Draw it" mode: click-drag a box directly on the page image to
add a region the layout engine missed, as typed/OCRed text or as a new
figure (add_manual_region, POST /api/magazine/{doc_id}/pages/{page}/regions).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Split-screen review: the rendered page (left) with a numbered box per
detected block, the extracted text (right) with matching numbers.
Clicking either side highlights the other. Checkboxes moved from a flat
list into this layout without changing their semantics (still gate
what gets narrated).
"OCR again": reocr_page() forces one page through the configured
vision engine again, replacing only that page's blocks -- deliberately
does NOT call build_articles(), so an already-approved article is not
silently re-grouped. A magazine-level option re-reads every page and
re-runs detection (with a confirmation, since that does reset article
status). extract_pages() gained force_vision_ocr to skip its quality
gate for this case.
Building the numbered-box review immediately did what it was built to
do -- surfaced a real bug on live production data (real magazine, real
scan, not a synthetic fixture): a photo caption at 49.9% of the page
width sat in the gutter between two genuinely separate columns and
merged them into one, so every block on the page came back column=0
and reading order fell back to pure top-to-bottom, interleaving the
two columns' sentences. Fixed two compounding causes: column boundary
detection now excludes blocks much wider than the page's median width,
and detect_columns' gutter threshold (3.5%) was itself narrower than
some real magazine gutters (~2.1% measured) -- lowered to 1.5%.
10 new/updated tests, including one built directly from the real
measured page-17 block coordinates that exposed the bug.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Magazine P1:
- Images shown automatically during playback (assign_images_to_articles
links each figure to the article whose column it falls in; the client
shows the last image reached in reading order as playback advances).
- "Play all articles" queues every approved article and advances
automatically at the end of each rather than stopping.
- Word-level highlighting, explicitly documented as an ESTIMATE (no
forced-alignment/timestamp source exists): each word's share of the
sentence's real measured audio duration is proportional to its
character length. A checkbox turns it off; sentence highlighting
(measured, not estimated) is unaffected.
- Driving mode: full-screen overlay, two oversized controls, reuses
whatever is already playing/queued.
- Sidebars/pull-out boxes excluded from narration automatically
(mark_sidebars: an anomalously narrow column vs. the page's regular
grid). Fixed a real defect found while building this: a full-width
headline was left in the column-width analysis and could bridge two
genuinely separate columns into one, hiding a real sidebar column
entirely -- caught by a failing test before the fix.
Second OCR engine + fallback ladder:
- KDL Frontier Parser nano wired in via vision_ocr_page_chain (ordered
engine list, first usable result wins, tesseract last resort).
Confirmed live: ~3x faster than Baidu (~12s vs ~40s/page) but returns
plain text with NO coordinate markers, unlike Baidu's labeled regions.
_whole_page_region treats the whole response as one ambiguous
full-page block (gated by ocr_quality_score) rather than discarding
usable text just because it lacks layout structure.
OCR engines page + a real UI bug:
- The OCR & Document Parsing page had no .s-engines-page[data-page="ocr"]
markup at all -- renderLocalContainers already supported an OCR grid
generically, but $('dc-grid-ocr') was always null, so the card grid
silently never rendered. Added an Active OCR Engine panel mirroring
Active Language Model exactly, plus the Local grid -- both configured
engines now appear as ready-made cards, no "Add Custom" dialog needed.
- Fixed the audiobook pipeline's "Previous/Next" nav bar bleeding into
every .section-head in the document (confirmed live on the Engines
page: a "Cast Audiobook" button and "STEP 1/8" hint that made no sense
there). Changed from a blocklist (naming each unrelated section as
noticed) to an allowlist of the pipeline's three actual destinations.
Verified live end-to-end: play-all queue building and manual advance,
word-span markup, driving-mode toggle, OCR page showing both engines as
local cards, zero stray nav bars on the Engines page, no console errors.
20 new backend tests (156 total in the magazine module).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A page needing OCR (or whose already-extracted text -- native or
tesseract alike -- scores as mostly garbage via ocr_quality_score) is
now escalated to a connected vision OCR engine before falling back to
tesseract. Falls back automatically on any failure (unreachable engine,
timeout, unparseable output); never fatal to ingestion.
Verified live against ocr-baidu-unlimited-ocr on a real 28-page scanned
c't magazine excerpt (Epson ES-580W, no scanner OCR): every garbled
article title tesseract produced ("K | soll Arbeit abnehmen", ". |",
"@)X", "Owns mee") was replaced by the real headline ("1. Irrtum: KI
ist ein allgemeiner Jobkiller", etc.), and furniture detection jumped
from 44 to 151 blocks because the engine labels running headers/
footers/page numbers directly instead of needing fuzzy post-hoc
matching. 26 of 28 pages used the vision path; the remaining 2 fell
back to tesseract automatically.
Also fixes two real defects found investigating the same real-world
data, both regression-tested against the actual garbled strings
observed:
- Exact-match repeated-line detection (running header/footer) silently
failed once tesseract OCR'd the same line differently each time --
cluster_running_furniture fuzzy-matches near-edge blocks instead.
- ocr_page() built one Block per OCR text-LINE rather than per
paragraph, so a single anomalous line (a chart number, scan noise, a
wrapped headline continuation) could independently trigger title
classification -- now grouped by paragraph like the native PyMuPDF
path, using the paragraph's median line height rather than its total
bbox height for font size.
A vision-engine label (footer/title/page_number/caption) is trusted
outright and skips classify_blocks' font-size heuristics entirely,
since the engine gives no reliable font metric; an ambiguous "text"
region gets a synthesized estimate from its paragraph area and length
(estimate_font_size_from_area) and still goes through normal
classification.
Off unless vision_ocr_url/vision_ocr_model are configured
(vision_ocr_enabled setting, default on if reachable); per-request
override via the ingest endpoint body. Default engine URL uses
host.docker.internal, confirmed live -- the OCR container runs on its
own Docker network, not dgx_net, so a container-name default would not
resolve from the app container.
28 new tests (136 total in the module). ocr_quality_score is
deliberately coarse and documented as such: it catches text that is
mostly symbol noise, not a plausible single-word misread inside an
otherwise normal sentence (confirmed live: a misread "KI" -> "K |"
still scored 0.81) -- that class of error needs a dictionary or
language-model check, out of scope here.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Studio -> Magazines: library, contents tree grouped by section with
confidence/pages/listening-time/audio status, the Stage D review
screen (tick/untick layout blocks, correct title/author, approve or
reject), a reflowed article view with click-to-seek sentence
highlighting, a linked original-page view showing the source bounding
box, and a podcast panel (2-3 host pickers, script generation,
turn-by-turn playback).
Wired into the existing lazy-module system (module-manifest.json,
loader.js, nav.js) rather than a separate app. Reuses /api/tts-preview
for both narration and podcast turns -- no new audio pipeline.
Fixes found building this:
- The pane toggle used role="tab" without the full ARIA tabs pattern
(aria-selected, roving tabindex, arrow-key nav) -- fails WCAG 2.2 AA
and hid the buttons from role-based tooling. Replaced with a button
group + aria-pressed, matching the plain-click interaction it
actually has.
- Print layout cues ("(continued on page 8)") were being sent to TTS
verbatim -- narration would have spoken the parenthetical aloud.
Stripped from the narration copy only; layout.json keeps the
original for the review screen.
- has_podcast now comes from the articles listing so opening an
article without a generated script skips a doomed GET instead of a
routine 404 on every open.
Verified against the real API contract with Playwright (library,
review pre-selection, approval, sentence highlighting + PDF-page
linking, podcast approval gate, zero console errors across the flow)
and manually end-to-end against a real ingested fixture, screenshots
included. Full suite: 94 Python tests, 33/34 desktop Playwright tests
(one unrelated pre-existing flaky DOM-count assertion).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Magazines break the audiobook pipeline's assumptions: several unrelated
articles share a page, columns run in parallel, adverts and sidebars
interrupt body copy, and a feature routinely jumps twenty pages to its
tail. Narrating page-by-page walks out of an article mid-sentence and
into an advert, so the canonical object is the article, not the page.
core/magazine.py reconstructs articles from a flat PDF: PyMuPDF extracts
blocks with bboxes and font metrics, tesseract supplies a TSV fallback so
scanned pages join the same column grouping, blocks are classified and
ordered column-major, and articles are anchored by the parsed table of
contents with "continued on page N" tails linked back.
routes/magazine.py adds 13 endpoints. Ingestion is a restart-safe
background job. Nothing is narrated until a human approves the
boundaries; podcast generation refuses unapproved articles.
Podcast mode renders an approved article as a 2-3 host discussion using
the configured LLM and voices the user picks, alongside narration rather
than instead of it.
Verified end-to-end over HTTP on a synthetic 8-page issue: the feature
came back as pages [3, 8] with its jump tail absorbed and the advert page
excluded, and an image-only rendering recovered all 47 blocks via OCR.
44 new tests; full suite 90 passing.
The Studio UI is not built — this is reachable only via the API.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- Deterministic hash-based character portraits and reference sheets
- Implemented strict early Renaissance aesthetic controls and role-specific headwear rules
- Added audiobook batch export scripts and progress tooling
- Verified book profiles with a new 'confirmed' state tracking mechanism
- Updated Pillow dependency for robust image handling
- Configured .gitignore for audiobook and audio cache exports
Fish-Speech clones from the app's own saved WAV library but was missing from
the set of backends populated with those voices, so the UI probed the Fish
server for a voice-listing endpoint it does not have, showed "Fetched 0
voices" and left the engine unusable despite being healthy. It now reports
the full library (218 voices).
Also adds a Fish-only panel listing all 49 documented emotion tags with
click-to-insert and a multi-emotion example, and corrects the backend's
advertised capabilities: it was flagged style_aware with "emotion markers are
honoured per request", but measurement shows a reference clip's in-context
prosody overwhelms inline tags (10x loudness spread across emotions without a
reference, 1.2x with one). It is now described by what it does well:
deterministic, byte-reproducible cloning.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Fish-Speech emotion tags were silently ignored on non-English books: per-line
emotions are LLM-generated in the book's own language, but Fish-Speech only
recognizes English [tag] markers, and a double-tagging bug was stacking a
broken server-derived tag on top of the client's own. Added a DE->EN
translation table and removed the double-tagging. Also wires the existing
book-profile context and race_species field into character portrait prompts
(previously only used for voice design), adds a recast-until-threshold loop
for casting, and adds backend-aware emotion quick-picks to Read Aloud, Try a
Voice, and Conversation.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
For a designed voice the instruct prompt is the voice's identity — the TTS
engine reproduces the voice from that text alone. The only copy saved was the
`note` display summary, clipped to 240 characters, which left 43 of 73 voices
cut off mid-sentence. Save the complete prompt in its own field so the engine
can register a voice from the whole description.
Existing voices keep working from the clipped copy (it still carries gender,
accent and timbre) and pick up the full text when next redesigned.
Pairs with the engine-side fix in tts-dgx-spark-faster-qwen3-tts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Voice consistency:
- Read back each voice's pinned seed (Seed Finder / Batch Seeds) on every
generation. The seed was saved to voice metadata but only ever read by the
Seed Finder's own benchmark path, so all per-voice seed pinning was inert.
- Stop coercing the "voice_design_playback" stability profile back to
"voice_clone". The pseudo-backend key isn't a real routing target, so the
backend-name normalizer silently rewrote it — reintroducing the hardcoded
seed:0 that profile exists to avoid, overriding every per-voice pin.
- Apply the accent clause on every line, not just at voice-creation time,
and reorder the instruct so emotion leads and accent trails (Qwen3-TTS
doesn't reliably follow multiple conflicting instructions).
- Pass an explicit language to Voice Design instead of leaving it on "Auto".
Audio effects:
- Add a limiter after compressor makeup gain. Makeup gain pushed peaks to
~1.9, and the final hard clip turned that into broadband distortion that
swamped the rest of the chain.
- Cascade highpass/lowpass 3 stages each (~18 dB/octave). Single-pole
filters were too gentle to band-limit speech audibly.
- Add a Bandpass control and wire it into the Telephone/Radio presets —
compression alone never sounded like a phone; band-limiting is the
defining trait.
Persona / Try It Out:
- Disable "Apply character persona" with an explanatory tooltip when the
voice has no persona saved, and error clearly server-side instead of
silently no-op'ing. Persona is typed manually per voice, never auto-filled.
- Stop dropping applyPersona in the chunked generation path (>200 chars).
- Populate the Voice Design dropdown from the user's own library rather than
filtering the engine's discovery list, which never contains custom voices.
Navigation and library:
- Use pushState instead of replaceState so browser Back/Forward step through
in-app navigation instead of leaving the app entirely.
- Show real dialogue line counts in the character sidebar instead of the
capped reference-quote count (which showed a misleading uniform "12").
Also fixes a crash in /api/transcribe-bytes that referenced an undefined
source_id in its cleanup path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Introduces the new Studio section (Source -> Characters -> Voices ->
Perform & Export) that reuses the existing Read Aloud/Library/Script
Rehearsal code via DOM reparenting instead of duplicating it, and rolls up
a long tail of bugs found while producing a real audiobook through it:
umlaut-eating name sanitizers, a voice picker that mispositioned itself and
capped results at 60, PDF pagination silently breaking on trimmed \f
markers, a race letting stale audio keep playing after a new line was
clicked, an alias-overlap bug that could silently redirect a voice/image
save onto the wrong character, voice design failing outright during brief
TTS backend restarts instead of retrying, sparse cast entries defaulting to
English/wrong gender, and a reassigned voice never reaching an already-open
Stage session or invalidating its cached audio. Also adds a persistent
per-line audio cache, audiobook export browsing/download, and an inline
voice-design prompt editor. Full details in CHANGELOG.md.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Same pattern as the Casting audiobook Prompt button: a collapsible
textarea with the full extraction prompt (fields + sources/citation
schema), named-preset save/load/delete, and live persistence so edits
apply from the next passage onward. Server-side, _charsheets_prepare()
now accepts a character_sheets_prompt override (shared by the blocking
and streaming endpoints), following the same pattern already used for
the audiobook attribution prompt override.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two gaps observed in real casting output:
1. A mid-quote " - " (em-dash pause) was being read as the end of the
quotation, handing the rest of the line to a wrong/new speaker.
2. Narration naming whose voice is about to speak (e.g. "Marcians
Stimme wirkte nicht mehr so fest") wasn't being used to resolve the
following unattributed line away from 'Unknown'.
Added as rules 11/12 to the client-side AB_DEFAULT_PROMPT (the prompt
actually sent to the LLM) and to normalizeCastingPrompt()'s migration
so already-saved custom prompts pick them up automatically, same
mechanism as the earlier Doppelpunkt-Regel upgrade. Also added
equivalent rules to the server-side fallback prompt in
routes/conversation.py for completeness, though the client-sent prompt
is what's actually used in normal operation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Confirmed by directly probing a streaming response from vllm-0.23.1rc1:
its delta objects carry reasoning text under "reasoning", not the more
common "reasoning_content" key every reasoning-display code path was
checking for. Added a shared _reasoning_text() helper that checks both,
used by the Conversation Playground, audiobook casting's live view, and
character sheets.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Reasoning-model chain-of-thought is now separated from the spoken/displayed
reply on the server (handles both a dedicated reasoning_content field and
inline <think> blocks, including chat templates that inject the opening tag
as a prompt prefix so it never appears in the stream). The client shows it
in a panel collapsed behind a "Thinking" chevron instead of dumping raw
reasoning text into the chat or speaking it aloud via TTS.
Also moved the noise-gate slider out of its own row and next to the mic
button, grouping it with Auto-stop/Hands-free/Live agent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Confirmed in production: a streaming attribution/character-sheets
request held _attribution_llm_lock for 11+ minutes, well past its
configured timeout, silently rejecting every subsequent passage as
"Attribution engine busy" and falling back to the non-streaming view -
looking exactly like live-thinking had stopped working, when actually
one earlier request never finished.
Root cause: requests' timeout= on a stream=True call only covers the
connect + first byte, not gaps between later body reads. If the LLM
backend goes silent mid-stream (connection left open, no more chunks),
the blocked socket recv() inside iter_lines() can hang indefinitely.
Since that's a native blocking call, not a Python-level yield point,
neither an in-loop wall-clock check nor GeneratorExit from a
disconnected client can interrupt it - both only take effect at the
next bytecode boundary, which never arrives while blocked in the C
extension.
Added _watchdog_close: a daemon thread that force-closes the upstream
connection if the wrapped block hasn't finished within the configured
timeout. Closing the socket from another thread makes the blocked
recv() raise, unblocking the generator so its normal except/finally
cleanup (including releasing the lock) actually runs. Verified against
the live LLM backend post-restart: stream completes normally with
real token-by-token deltas, and the lock is confirmed free immediately
after.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Character Sheets generation only ever showed a progress bar - no
visible reading/thinking/filling-out, unlike the casting flow which
already streams the LLM's output live. Refactored
/api/character-sheets into shared _charsheets_prepare/_charsheets_parse
helpers (same split used for attribution) and added
/api/character-sheets/stream, proxying the LLM's SSE stream through
the same shared lock used by the other attribution endpoints.
Client: new csGenerateStream (mirrors audiobookAttributeStream) tries
the streaming endpoint first per passage, updating a new "Live output"
panel in the progress dialog with the raw JSON answer as it's written
- itself the "watch it fill out the sheet" experience, since there's
no separate reasoning channel worth hiding it behind here. Falls back
to the blocking endpoint on any stream failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Table view was rendering as stacked blocks instead of columns: rows
reused .lib-char-card for its event wiring, but that class's
display:flex;flex-direction:column turned every <tr> into a flex
column. Reset to display:table-row and stripped the leaked-in card
chrome.
Split /api/character-generate-prompts into four independent per-field
LLM calls (from two paired calls) and added a `fields` filter, so the
UI can offer one Generate button per prompt box instead of a single
button that always regenerated all four - cheaper, and further
shrinks each response to reduce truncation risk.
Added a Sort dropdown (Role/Alphabet/Lines/Gender/Voice assigned) to
the Characters/Cast list, persisted like the Cards/Table toggle.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Table view: one row per character (avatar, name, sex, line count, voice
language, alignment, voice, tags, prompt-availability checks), toggled
next to the card grid and persisted.
Bulk voice auto-assign: checkbox per character + "Auto-assign selected"
per production, sequential so later picks see what earlier ones just
took (avoids duplicate voice assignments).
Character tags (auto-set to the book of origin) are now visible on
cards - the field always existed, cards just never rendered it, so a
character recurring across books had no visible link between records.
Detail fields (Backstory, Motivation, etc.) now show small numbered
links to their exact source citation when the sheet has one, instead
of making the reader search the full "Quellen im Text" list.
Added gender as an actual extracted character-sheet field - the UI
already had a gender icon but the LLM was never asked for the value.
Fixed SillyTavern/Concept Art prompts still coming back empty despite
the earlier token-budget increase: they're the last two fields in one
JSON object, so truncation always cost the same two regardless of the
ceiling. Split into two independent, concurrent LLM calls instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The streaming attribution endpoint decoded the LLM's SSE response with
requests' guessed encoding (Latin-1 fallback when no charset is declared),
mangling every German umlaut. Forced UTF-8 explicitly.
The "LLM Thinking" pane duplicated the passage text for models that
ignore the <think> instruction and stream straight into JSON - it now
only shows real reasoning when present, and otherwise labels raw output
honestly instead of passing it off as thinking.
Also fixed three UI bugs found while testing a live multi-hour cast:
- A-/A+ font buttons had no effect (a hardcoded font-size on .ab-cv-row
always overrode the CSS variable they set).
- Typing a name + Enter in the "Assign to" popup (and drag-to-assign,
which reuses it) silently did nothing after the first cast/recast run
in a session - the popup is a page-lifetime singleton but its input
handlers closed over the first run's now-stale assignName/closePopup.
Every popup open now repoints them at the current run.
- "Split text to Unknown Speaker" split at the wrong spot when the
selected phrase repeated earlier in the same paragraph (indexOf found
the first occurrence, not the dragged one). Now uses the exact DOM
range offset instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The static-asset caching middleware used BaseHTTPMiddleware, which has a
known Starlette bug: a client disconnecting mid-StreamingResponse (the new
live-attribution SSE stream hitting its idle timeout) raced its internal
task group and raised "RuntimeError: No response returned", crashing that
request. Rewritten as plain ASGI middleware that only touches headers via
the raw send callable, removing the race.
Also found the real cause of the casting timeouts/405s: the streaming
attribution endpoint had its own lock instead of sharing the one the
blocking endpoint already used to serialize on the LLM's single slot -
letting a stream call and its own blocking fallback fire concurrently,
exactly the ghost-request pile-up that lock was built to prevent. Unified
onto one lock and added server-side logging for stream failures.
The "LLM Thinking" pane now shows the model's actual <think> reasoning
instead of the in-progress JSON answer echoed back at the user.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The LLM Reading card splits into thinking-stream (left) and passage
(right), fed by a new SSE endpoint that shares prompt-building and
parsing with the blocking one and falls back to it on any stream
failure (inactivity timeout, not overall). The deterministic resolver
no longer invents speakers from scenery nouns (PLATZ/GESICHTER/
KLEINIGKEIT) — person-noun whitelist plus a clause-subject pattern —
and gains the impersonal post-quote formula and strict two-person
alternation with colon/page/window guards. All reported failure cases
verified against the exact book sentences.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The saved casting prompt was the 2nd-quality verification prompt, so
first-pass attribution ran with the wrong job description and none of
the deduction rules — and the client-side prompt migration had no
anchor to upgrade. The server now appends the rules to any prompt
lacking them, and the saved prompt was reset to the default (backed up
to config/audiobook_prompt.backup.txt). Also: single-pass combined name
regex (a shorter alias could match inside a longer name's data-name
attribute and leak raw style="..." into the feed), stopword filter so a
comma-split alias like "Die, die den Vampir verließ" can't underline
every article, heading-like narration renders bold/centered, and A-/A+
font controls in the casting toolbar sharing the Rehearser Stage scale.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Characters/Cast gains a "Casting" button back to the active casting
session, the pipeline stepper renders on the Library section, and the
stepper's Cast Characters stop navigates instead of side-effect-running
sheet generation. Prompt generation: 4096-token budget (1600 truncated
the four-prompt JSON so two fields silently arrived empty), truncated
answers salvage completed fields, all-empty responses fail loudly, and
partial results name the missing prompts. The PDF-extraction progress
pill is enlarged and vertically centered.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The per-word <span> wrapping behind "click any word to assign" created
~100k DOM nodes at book scale and froze the tab on every feed redraw;
replaced with native caretRangeFromPoint word detection plus a single
reused hover overlay — same UX, zero extra DOM. Export button gained a
2s re-entry guard (queued clicks during a freeze fired as a download
burst) and now delivers one zip: the cast script in Markdown plus a
sheet per character. Character detail view gains a Generation Prompts
section — four fold-out copy boxes (Voice Design, Character Image,
SillyTavern card, Concept Art sheet) filled by one LLM call over the
full profile via the new /api/character-generate-prompts endpoint.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Add a 6-stage pipeline stepper (Source -> Cast Audiobook -> Cast Characters
-> Script Rehearser -> Generate MP3s -> Audiobook) with direct, non-destructive
jumps between stages and a prominent guided-tour look
- Split PDF import into an explicit "load" then "Extract Text" step, with
in-browser OCR (Tesseract.js, vendored) to recover chapter headlines baked
into a PDF as images instead of real text
- Fix casting feed silently merging pages after leaving/returning: segments
now carry their own page number instead of re-guessing it from text
- Fix excessive "Unknown" speaker attribution: restore the attribution LLM's
output token budget, which had been cut roughly in half and was truncating
dialogue-dense passages
- Fix Theater Play library cards failing to open (dead pre-migration
IndexedDB API calls, missing section navigation)
- Fix bulk "Set tag" wiping a voice's existing tags instead of adding to them
- Start merging Casting's feed with Script Rehearser's Stage UI: collapsible
character sidebar, shared "paper" page styling, inline text editing
- Fix a performance regression from that merge (per-row listeners on every
redraw) by moving to event delegation
- Various layout/clutter fixes: hide reader chrome until a document is
loaded, collapse secondary settings by default, fix overlapping toolbar
icons, fix duplicate "opening" notifications
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Exposes the `speed` parameter (0.5–2.0) on the TTS `/v1/audio/speech`
request so audio is generated at the target tempo natively via the
faster-qwen3-tts backend rather than post-processing.
- Backend: `/api/tts-preview` now extracts and forwards a `speed`
override (clamped 0.1–4.0) through the same extra-params mechanism
already used for seed/temperature; backends that reject it fall back
cleanly via `_post_tts_with_fallback`.
- Try it out: "Native Speed" number input (0.5–2, step 0.05) added to
the text card; value persists in localStorage per browser; passed as
`extra` through `createTtsAudioSource` and `generateChunkedTts`.
- Read aloud: "Native Speed" control added to the generation controls
row alongside seed/temperature; included in `readerGenParams()` and
saved/restored with library documents (each book tracks its own speed).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a persistent, book-scoped Character Library (new Characters section,
IndexedDB) that auto-fills from Character-sheet analysis with editable cards.
Enrich extraction with six narrative fields (backstory, relationships,
motivation, fears, mannerisms, voice/speech) plus the greyscale Good↔Evil
alignment bar, arc arrow, and 5-area Deep Analysis. Add ⋯ separators between
non-contiguous passages in Recast unknown, and harden dialogue attribution
against hallucination with same-language emotion tags.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
