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.
Migration 0004 — authors + author_revisions + stories.author_id +
stories.author_revision_id + stories.cross_story_memory +
author_corpus. Soul versioning built in from day one per cobb's
locked decisions:
- authors.id immutable identity (slug + display_name + tagline + model)
- author_revisions tracks each soul revision with n monotonic
- Partial unique index 'idx_author_revisions_current' enforces
exactly one is_current=true per author
- stories.author_revision_id pins to the exact soul used at gen
time (so 'this was the Orson Black active when chapter 8 was
written' is always recoverable)
- author_corpus tracks 'authored' + 'read' relationships for the
v0.3 cross-story memory toggle
skald-core::authors module — CRUD: get_by_slug,
get_with_current_revision, get_current_revision, get_revision,
create_or_get (idempotent), add_revision (transactional, demotes
prior is_current=true), assign_to_story (also touches
author_corpus).
Web v0.1 forms (the second feedback bucket — 'no way to make new
stories', 'no options for sequels'): handlers + form panels +
POST routes for /stories/new and /stories/:id/continue. Both
create a story stub with status='seed'; actual generation will be
fired by 'skald continue' (next commit) walking seed rows.
Norse visual revamp + mobile collapse deferred — vetting full gen
is the priority per cobb's 'green light for v0.3'. Coming back to
the aesthetic after the pipeline works end-to-end against a real
Orson Black-authored Chapter 8 of Coast-Down.
Closes the TTS schema layer. The v0.2 render pipeline auto-runs an
audit chain after each chapter narration:
F5 render → narration_runs (succeeded)
→ ffmpeg chunk into ~30s windows
→ Whisper-large-v3 STT each chunk
→ word-level diff vs source chapter text
→ mismatches → narration_findings (kind=pronunciation|skip|insert)
→ ffmpeg silence/clip detect → narration_findings (kind=glitch)
→ (optional) Gemini Flash audio review pass
→ narration_findings (kind=prosody|tone)
→ unresolved crits trigger automatic re-roll with new seed
Distinct from audit_findings: that table is canon/continuity at the
text layer, populated by the third-Opus canon-audit pass.
narration_findings is audio-quality only, populated by detectors
that consume the rendered WAV.
The 'detector' field captures which model produced the finding so
we can tune thresholds per detector when one over- or under-flags.
cobb's audio agent intuition was right: STT-and-diff catches the
'name came out wrong' case airtight, and a separate audio-native
LLM call catches the subtler 'this sentence sounded weird' cases
Whisper can't see.
TTS layer landed as schema-only — synthesis pipeline ships in v0.2.
Putting the tables in v0.1 means imports already carry the right
shape; we won't need a 'migrate every existing story' pass later.
Decisions locked 2026-05-13:
- Engine: F5-TTS (best 8GB FOSS option, mid-2026 SOTA)
- Default voice source: LJ Speech (Linda Johnson, PD released
specifically for TTS training — airtight for sharing/uploading
generated audio. The 'AI-consent-released' license posture is
the difference between 'should be fine' and 'definitely fine.')
- Variety voices: Hi-Fi TTS speaker IDs (Apache 2.0, same consent
shape). LibriVox is optional but never default.
- Pronunciation overrides DB layer (story-scoped + global) to fix
proper-noun mispronunciation — the actual TTS-quality gap on
Cobb's bar of 'must not wake me up.' Pre-pass with Opus extracts
proper nouns + IPA, operator verifies, table caches forever.
Tables:
- voices — name, license, reference_path/text, sample_rate, default flag
- pronunciation_overrides — story-scoped or global, IPA/arpabet
- narration_runs — TTS audit trail mirroring generation_runs
- stories.preferred_voice_id FK
Unique constraints:
- one default voice (partial index)
- one row per (story, word) override
- one global row per word
Skald is a generic story-writer. The database is the product; the
binary is the tooling. Everything story-specific lives in rows, not
in code. cwho's monorepo + binary-per-role pattern transplanted to
this domain.
What this commit ships:
- Cargo workspace (resolver=3, edition 2024): skald-core (lib) +
skald (bin)
- Migration 0001: stories, characters, canon_facts, chapters,
chapter_summaries, passages (vector(1536)), generation_runs,
audit_findings, tags. pgvector + pg_trgm extensions. ivfflat
index deferred until we have data (post-import the first ~1k
passages and add the index).
- skald-core::ingest — markdown parser for the cwho/coast-down shape:
'# Title' → '## Chapter N — date' headings → '# Continuity Bible'
section with character roster (real + fictional sub-sections) +
setting / mystery / historical / liberty / hook sub-sections.
Decomposed into structured rows; original bullet body preserved
in key_facts/body fields for fidelity. 6 unit tests cover the
shape.
- skald-core::db — Postgres connection pool + migration runner.
- skald-core::models — row types via sqlx::FromRow.
- skald binary — clap CLI: 'serve' (http + migrations) and
'import-markdown' (one-shot ingest).
- Dockerfile — multi-stage: rust:1.95-bookworm builder, pgvector/
pgvector:pg17 runtime, tini under PID 1, custom entrypoint.sh
that boots embedded postgres then execs skald serve.
- compose.yml — singleton container, postgres data in volume,
story corpus mounted read-only at /seed.
Decisions locked 2026-05-13:
1. DB in same container 'till we have a real working tool' (cobb)
2. postgres+pgvector (NOT sqlite) — keeps semantic-search story
3. Network-not-socket connection (postgresql://localhost:5432) from
day one so future split is config-only, not code-rewrite
Not yet wired:
- Web UI
- clawdforge calls (gen → cleanup → canon-audit pipeline)
- Embedding pass
- TTS sidecar
