the host bind paths + LAN host pins replaced with env defaults. Repository URLs
→ git.sulkta.com. Audit-changelog scaffolding stripped from inline comments
(technical reasoning preserved). README sheds marketing scaffolding. AI-speak
in load-bearing prompts/SOULs left alone — that IS the product.
One chapter in a dedup run had the model emit the chapter twice
(18.7k -> 32.6k chars). A surgical dedup only nudges length, so
the pass now rejects any output wildly off the input length —
marks that chapter's run failed, leaves the prose untouched, and
moves on rather than silently corrupting it. New --chapter flag
narrows a run to one chapter so a skipped one can be retried.
Adds `skald dedup --story <id>`: reads a story's most recent
prose-audit findings and walks every chapter, handing the author
the chapter prose + the findings with instructions to rephrase
ONLY the flagged repetitions (each recurrence made distinct) and
fix flagged continuity errors — everything else stays verbatim.
A surgical dedup, not a rewrite. Overwrites body_md, clears
body_md_tts so the chapter is re-prepped before narration. High
effort (prose-craft). Migration 0011 adds the 'dedup' run kind.
Completes the QC loop: audit (find) -> dedup (fix) -> re-audit.
A max-effort pass over a whole-book input (quarter-million chars)
runs an impractical hour-plus and was timing out. Finding repeated
passages is comparison work, not deep reasoning — high effort does
it well in minutes at a fraction of the cost.
The prose audit reads a whole story at max effort and genuinely
runs past 1800s on a 50k-word book — it was timing out. Bump the
audit pass server-side timeout to 3600s and the shared reqwest
client ceiling to 7200s (matching clawdforge's new cap). Other
passes keep their own shorter timeout_secs.
Adds `skald audit --story <id>`: a whole-story QC pass that reads
every chapter end to end and flags repetition, template tics,
self-restatement and continuity drift — the gate before a story
goes to narration, where repetition a silent reader skims is
glaring read aloud. Runs at max effort (real reasoning work,
worth the spend); findings land in audit_findings and print.
Also hardens the gen + cleanup directives to hunt repetition at
the source: re-phrase recurring motifs fresh, no stacked template
anaphora, dialogue echoed verbatim at most once.
Migration 0010: 'prose_audit' generation_runs.kind, 'repetition'
audit_findings.area.
Adds GET /stories/{id}/listen: one <audio> element over a story's
stitched audiobook file plus a clickable chapter list. Clicking a
chapter seeks; the chapter under the playhead highlights as it
plays. Chapter offsets are summed from each chapter's latest
succeeded narration_run duration — the same order the file was
stitched. One small inline script, the web UI's first JS.
New stories.audiobook_path column (migration 0009) holds the
served path; the story page shows a "listen" action when set.
Two fixes:
- narrate_prep in single-voice mode (empty character roster) was
still handed the multi-voice directive, so the model invented
[voice:<slug>] tags from character names in the prose. The
narrate path neutralised them by falling back to the narrator,
but it was log spam and a leak of intent. Single-voice now gets
directive + house-system variants that forbid voice tags
outright, and the user-prompt task line matches.
- Every narrate run wrote a fresh ~80MB WAV and never reclaimed
the previous one, so re-renders piled up stale files. A
successful render now deletes the WAVs of prior renders of the
same chapter and nulls their output_path. Render history rows
are kept; only the dead file pointer is cleared. Best-effort —
cleanup failure never fails the render.
gen, cleanup, narrate_prep and rewrite drop from max to high
effort. Audit keeps max — it is the one pass doing real reasoning
(canon drift, timeline gaps, retcons) rather than prose-craft, so
it is worth the frontier spend. Prose-craft is "good enough" at
high. This also keeps the all-Opus skald pattern under the
$200/month claude -p cap landing next month.
New Forge::rewrite + PassKind::Rewrite. An author re-authors
existing chapter prose entirely in their voice — sentence rhythm,
word choice, paragraph shape all become theirs — while canon
(names, dates, places, events, order, technical facts) is preserved
exactly. Not editing; re-authoring. SystemMode::Replace, max effort.
skald rewrite --chapter <uuid> [--author slug] overwrites body_md
with the rewritten version. The pre-rewrite prose is stashed in the
new chapters.body_md_original column on first rewrite (migration
0008, idempotent) so the original is never lost. body_md_tts is
cleared — it was annotated against the old prose and must be
regenerated by a fresh prepare-narration.
prepare-narration gains --single-voice: skips the character speaker
roster so no [voice:X] dialogue tags are inserted, only beat
markers. Right for one-voice narration.
Migration 0008 also extends generation_runs.kind to allow 'rewrite'.
Engine dispatch was a binary kokoro-vs-everything-else split that
sent tortoise voices to F5. Now three branches:
kokoro_* → KOKORO_URL (the host default :7794)
tortoise_* → TORTOISE_URL (the host default :7795)
* → F5_TTS_URL (the host default :7792)
substitute_speaker_voices also runs for tortoise voices, so
multi-voice [voice:slug] tags survive in the prose. Tortoise voices
without a per-character mapping fall back to the narrator voice
server-side (tortoise_server.py logs the fallback).
The python FastAPI sidecars have lived ad-hoc at /srv/appdata/
<engine>/build/ on the host without version control. Bringing them into
the skald repo so the engine code travels with the cross-engine
routing it depends on.
This commit lands the VANILLA version of each engine on main:
engines/f5-tts/ SWivid F5-TTS (CC-BY-NC weights flagged)
engines/kokoro/ hexgrad Kokoro-82M (Apache 2.0 top to bottom)
engines/tortoise/ neonbjb Tortoise-TTS (Apache 2.0 top to bottom)
Engine-specific kludges (question doubling, GPU coordination,
pause-duration tuning) get layered on engine/* branches per the
README. Main stays the safe-to-read baseline.
skald narrate pre-processes body_md_tts with word-boundary regex
substitutions from pronunciation_overrides where phoneme_format =
'respelling'. Story-scoped overrides win over global; longer words
substitute first (so 'Bryukhanov' wins over a hypothetical 'Bry'
override). Case is preserved on the first letter so 'Pripyat' at
sentence start stays capitalised after respelling to 'Prip-yat'.
Seeded Coast-Down with 52 entries: Russian/Ukrainian surnames
(Dyatlov, Akimov, Bryukhanov, Stolyarchuk, Yuvchenko, Khmel, etc.),
first names (Pyotr, Lyudmilla, Anatoly, Vasily, Sasha, Aleksandr,
Leonid), patronymics (Stepanovich, Fyodorovich, Mykolaivna,
Hryhorivna), and places (Pripyat, Chernobyl, Kyiv, Kopachi,
Lubyanka). Plus the operational acronyms NIKIET, RBMK, AZ-5, SIUR,
SIUB, ChNPP, MSCh.
Other phoneme_format values (ipa, arpabet) are no-ops for now —
Kokoro's misaki tokenizer doesn't expose a stable lexicon-injection
API across the HTTP boundary in v0.1. Future: pass IPA forms in a
new server-side request field and inject into the pipeline's g2p
lexicon for more accurate phonetics.
Schema: characters.voice_id + characters.slug (migration 0007).
voice_id is FK to voices(id); slug is the stable lowercase token
the narrate_prep pass uses inside [voice:slug]...[/voice].
Forge::narrate_prep takes &[CharacterSpeaker]. System prompt
expanded to instruct the author to wrap dialogue lines in voice
tags based on a roster supplied in the user prompt (slug + name +
short hint from key_facts). Unattributed dialogue stays unwrapped
and inherits the narrator voice.
skald narrate substitutes [voice:<character-slug>] →
[voice:<kokoro-voice-name>] right before sending to Kokoro, using
characters.voice_id JOIN voices.reference_path as the map. Slugs
with no voice or no character row fall back to the narrator voice
defensively (logged as warn).
kokoro_server.py v0.4: splitter recognises [voice:X]...[/voice]
blocks at the paragraph level. Each text node carries an optional
voice attribution; renderer feeds it to Kokoro per-segment. Outside
voice blocks the request's default voice is used. voices_used is
reported back so callers can verify multi-voice actually ran.
Only kokoro-routed renders pre-process voice tags; F5 paths leave
the tags in place (F5 multi-voice not implemented). Defensive
fallback: orphan/unclosed [/voice] markers are silently absorbed
rather than failing the render.
Caught when redeploying after the 0006 patch: the live DB had
migration 5 stamped with a stale checksum + the column already
present, so neither re-apply nor checksum-only-fix worked cleanly.
Making 0005 idempotent fixes both paths.
Migration 0005 added the chapters.body_md_tts column but missed
this check constraint update — caught at runtime when
prepare-narration tried to insert kind='narrate_prep'.
Postgres doesn't ALTER CHECK in place; we drop + re-add.
Two new things working together:
1. Migration 0005 adds chapters.body_md_tts (nullable). Narrate path
prefers it over body_md when present — that's the annotated-for-
audiobook variant. Falls back to body_md if not set.
2. New Forge::narrate_prep pass: author (or House) annotates prose
with [breath] / [pause:Xs] / [scene] beat markers AND occasional
humanizing narrator stumbles (em-dash repetition, self-correction,
hesitation — sparingly, 1-3 per chapter). Apart from stumbles, the
prose is verbatim. Author voice threads through.
3. New CLI: 'skald prepare-narration --chapter <uuid> [--author slug]
[--overwrite]'. Records as generation_runs row kind=narrate_prep.
4. skald narrate now routes by voice.source — kokoro_* voices hit
KOKORO_URL (Apache 2.0 stack, audiobook-tuned with the v0.2 render-
and-stitch server), everything else hits F5_TTS_URL (voice-cloning
path). Voice DB row carries source as the dispatch key.
Why no new tag for narrator stumbles: em-dash repetition and self-
correction are just prose patterns Kokoro reads correctly because of
its punctuation cues. No new server-side machinery.
Caught on the 2026-05-13 Coast-Down 10-chapter Orson run: the LLM
labeled two chapters 'Chapter 1' instead of 9 and 10, and
ON CONFLICT (story_id, n) DO UPDATE silently overwrote them.
8 visible chapters from 10 successful gen+cleanup passes; 27k
words of work, ~6k buried. The audit caught the symptom but the
data damage was already done.
Fix:
- continue_story::run computes next_n from MAX(chapters.n) before
the batch loop; each iteration's authoritative n is next_n,
incremented after success.
- forge::generate + cleanup take chapter_n: Option<i32>. The gen
prompt is now 'Write Chapter N. Begin with: ## Chapter N — ...'
instead of the vague 'Write the next chapter.'
- We still parse_chapter() the LLM output but only to extract the
title; if the LLM-returned n disagrees with ours, we log a warn
and use the authoritative N at INSERT time.
The (story_id, n) unique constraint stays — it's now a defensive
catch for skald bugs, not the LLM's free-spirited numbering.
Chapter view now shows a narration card between title and prose
with three states:
- succeeded → HTML5 <audio> + voice + duration + download link
- running → 'rendering…' banner with relative start time
- none/failed → 'Render audio' POST button (spawns background
tokio task calling narrate::run)
ServeDir mounted at /audio serves WAVs from the f5-tts bind-mount
read-only. Range requests work, so 16-min chapters seek cleanly.
Deploy needs: compose mount /srv/appdata/f5-tts/audio:/audio:ro
on skald (already staged in /srv/appdata/skald/compose.yml on
the host).
forge.rs threads Effort::Max on gen + cleanup. Audit + summarize stay
default — they're structured-output / tool-shaped tasks where extended
thinking doesn't help. Bumps subprocess timeout from 600s to 1800s so
max-effort prose-craft has the wall clock it needs.
continue_story::run takes a chapter_count param; loops gen+cleanup per
chapter with each iteration's just-written prose appended to context.
Audit fires once at end against the combined batch vs parent canon.
Cap is 20 (~5h wall clock, ~$600 at max effort — beyond that is
operationally absurd).
CLI: 'skald continue --chapters N'. Web: numeric field on both new-
story and continue forms, 1..=20, defaults to 1.
Vendored clawdforge SDK refreshed for the Effort enum.
