v0.12.0

[0.12.0] - 2026-10-18

Generated-prompt quality pass across the deterministic templates and the LLM
rewrite, verified end-to-end against a local reasoning-model endpoint.

Added

• core/render.py: shared prompt primitives — scene_phrase/scene_noun
  (naturalized, slugline-free setting), per-shot seed_key, camera_sentence
  with a used-camera set for variety, soundscape_sentences (real audible cues
  from the script's action), and music_line (N/A unless a score is named).
• core/characters.py: capture appositive intros (DANI, a nervous junior reporter, ...) as character descriptions.
• core/llm.py: project_context + Rewriter (one cached per-project context
  call), a context= parameter threaded through messages/rewrite, a
  validate→repair loop (llm.fix_attempts), and sanitize to repair the
  format deviations small models make (inlined field labels, mis-placed dialogue
  language tags, dialogue stranded after non_diegetic_music, over-long audio
  fields, and cut times like At 3.87s normalized to MM:SS.mmm).
• core/llm.py: llm.disable_thinking (default true) sends
  chat_template_kwargs: {enable_thinking: false} so reasoning models (Qwen3
  etc. via koboldcpp) don't spend the whole token budget on a chain-of-thought
  preamble.
• tests: test_llm.py sanitizer/thinking-toggle coverage and
  test_quality.py::TestGeneratedPromptQuality (no raw sluglines, no timestamp
  on [Shot 1], camera variety, dialogue-free soundscape, N/A default music).

Changed

• core/prompts.py: stronger _COMMON/_R2V system prompts — naturalized
  setting, no raw slugline/transitions, camera variety, soundscape = sounds
  only, music N/A default, explicit output skeleton and dialogue format.
• core/fallback.py, core/r2v.py: build shots from the new render primitives
  (per-shot seeds, varied cameras, derived soundscape, music rule).
• core/config.py + config.toml: new llm.fix_attempts, llm.context,
  llm.disable_thinking options.
• __main__.py, app/web.py: construct llm.Rewriter (with character bible +
  context) when the LLM path is enabled.

Fixed

• core/characters.py: possessive cue names keep the apostrophe lowercase
  (WILL'S DATE -> Will's Date, not Will'S Date).

v0.11.1

[0.11.1] - 2026-10-17

Real-screenplay hardening: ran the full Big Fish feature (Fountain/FDX/PDF)
end-to-end and fixed the two issues it surfaced.

Added

• samples/Big_Fish/README.md: pointer for a locally-obtained Big Fish
  (2003) screenplay as an industry-standard pipeline test. The screenplay
  itself (© 2003 Columbia Pictures, not public domain) is git-ignored and
  not committed or shipped in release archives.

Changed

• parsers/pdf.py: extract with pypdf's layout mode (falling back to the
  default when unavailable) so hard-wrapped screenplay PDFs keep their paragraph
  breaks. Previously the whole script collapsed into one scene (B27).
• .gitignore: ignore build/ and dist/ (setuptools artifacts).
• Version 0.11.0 -> 0.11.1.

Fixed

• core/chunker.py: dialogue length is now measured cumulatively per cue,
  so a monologue written as many short lines (or reflowed PDF text) can no
  longer exceed the 10s cap; the cue is re-emitted per continuation piece with
  <scenetrans> crossings (B26).

v0.11.0

[0.11.0] - 2026-10-17

Added

• IMPROVEMENTS.md: Post-v1 roadmap (v2) consolidating the agreed v2
  direction — Project Manager (persisted, chunk-addressable projects with
  location/prop/continuity bibles), ComfyUI pipeline integration
  (draft/production workflows), Settings/Options page, pywebview desktop
  bundle, and a llama-server sidecar engine, plus prerequisites
  (schema_version, stable chunk IDs, core/h3_spec.py) and the later
  story/book → screenplay direction.
• next-steps.md: handoff notes for the next session (status snapshot, loose
  ends, blocking prerequisites, locked v2 decisions, verification commands).
• Joint (compound) speaker IDs (B14, guide §4.4): a cue like
  ELENA AND MARA (also &, /, +) now splits into individual bible entries
  and renders a compound ID — Elena and Mara (S1,S2) say <d>…</d>.
  (characters.split_joint, used by build_bible, chunker, fallback,
  ref_manifest); the validator now checks every ID in a compound group.
• pyproject.toml: [project.scripts] entry point (script2prompt), project
  URLs, readme, keywords, and Python 3.11–3.14 classifiers.
• .github/workflows/ci.yml: unittest matrix on Python 3.11–3.14, a
  zero-dependency import check, and an optional-extras job (pdf/web/dev + ruff).
• pipeline.write_pack(..., clean=True) and a CLI --clean flag: stale pack
  files (fixed export names + chunk_*.txt) are removed before writing.
• schema_version on persisted artifacts: manifest.json and
  reference_manifest.json now stamp pipeline.SCHEMA_VERSION (1). New
  pipeline.migrate_artifact / pipeline.load_artifact upgrade pre-versioning
  artifacts (missing field = version 0) through registered _MIGRATIONS and
  pass newer versions through untouched, so an older loader cannot corrupt a
  newer writer's output. First slice of the v2 persistence prerequisites.

Changed

• Dependency policy docs (AGENTS.md, PROJECT_PLAN.md §3.2) now state the
  locked v2 stance explicitly: zero-dep core is the default and a directive,
  not a fanatical rule; a 1.0 desktop bundle may add heavier optional
  extras while core stays dependency-free.
• parsers/plaintext.py: a slugline-looking line inside a dialogue block is now
  kept as dialogue instead of ending the scene (dialogue-state context).
• core/fallback.py: the unsupported-mode error now explains that build()
  handles T2VA/I2VA/FL2VA/L2VA and R2V is built by core/r2v.py (the stale
  "arrives in Milestone 3" message is gone).
• Docs: test-count references updated (124 → 149) in PROJECT_PLAN.md and
  TESTING.md; README.md documents --clean, the console entry point, and
  joint speakers.
• TESTING.md: .spmd is listed with .fountain (the Fountain parser), not as
  plain-text (B25).
• Version 0.10.0 -> 0.11.0 (second deep-audit bug-fix release).

Fixed

• validator._check_speakers: compound speaker groups ((S1,S2)) are now
  checked ID-by-ID against the bible rather than only the first ID.
• parsers/__init__.py: decode inputs as utf-8-sig so a leading UTF-8 BOM no
  longer corrupts the first line (B16).
• ref_manifest.propose: a wordless chunk with no slugline now synthesizes a
  placeholder environment subject instead of emitting an empty
  subject_definitions section that the validator rejects (B17).
• fallback._is_voiceover: normalize each parenthetical independently so
  MARA (V.O.) (CONT'D) (and (V.O.) (O.S.)) still render as a voiceover
  (B18).
• characters.py: joint action intros (MARA AND LEO (30s)) split into
  individual characters that share the traits, and all-caps mentions no longer
  swallow connectors (ELENA AND) (B19).
• characters.title: alphanumeric designators keep their casing (C3PO,
  R2D2) instead of becoming C3Po/R2D2-style mixed case (B20).
• parsers/fountain.py: a slugline-looking line inside a dialogue block is kept
  as dialogue (matching plaintext) (B21).
• config.load: invalid numeric env overrides (S2P_TEMPERATURE/S2P_TIMEOUT)
  raise a friendly error naming the variable (B22).
• pipeline.reference_manifest: returns an empty checklist for T2VA instead of
  raising no picture slots for mode: T2VA (B23).
• r2v.build: a blank style falls back to Live-action, cinematic rather
  than emitting uses a style (B24).

v0.10.0

[0.10.0] - 2026-10-17

Added

• core/render.py: shared prose helpers (clean, finish, soundscape_sentences)
  now used by both fallback and r2v (removes duplicated soundscape logic).
• core/llm.py: LLMError raised for unreachable endpoints, non-2xx responses,
  and malformed payloads; max_tokens/timeout/retries settings with
  exponential backoff on transient failures (5xx / network / timeout).
• config.toml [llm] max_tokens, timeout, retries keys; S2P_TEMPERATURE
  and S2P_TIMEOUT environment overrides in core/config.py.
• pyproject.toml dev extra (ruff) with [tool.ruff] lint config.
• LICENSE (GPL-3.0-or-later) with matching pyproject.toml metadata.
• install.sh: when the web extra is chosen it explains Streamlit's default
  telemetry collection and, if interactive, offers to write
  [browser] gatherUsageStats = false to ~/.streamlit/config.toml (preserving
  existing content, and flipping an existing gatherUsageStats = true).

Changed

• core/chunker.py: long action beats are split at sentence boundaries instead
  of being clamped to one chunk; a sub-5s trailing chunk merges into its
  predecessor when the pair still fits the 10s cap (the 5s floor now applies
  only to a chunk that stands alone).
• core/characters.py: title() preserves interior capitals (MARY-JANE ->
  Mary-Jane, O'BRIEN -> O'Brien) and the caps-name detector matches
  apostrophized/hyphenated names.
• core/ref_manifest.py: the environment subject uses the parsed slugline
  location/time ("Newsroom at night") instead of the raw INT. ... - NIGHT.
• core/h3_builder.py: falls back to templates only on llm.LLMError;
  unexpected exceptions from a custom rewrite now propagate.
• Version 0.9.1 -> 0.10.0.

Fixed

• core/llm.py: chat no longer surfaces raw HTTPError/JSONDecodeError,
  block silently on a hung endpoint, or leak the error response body.
• Web wizard: the "Write pack to ./out" button exported nothing because the
  generated project lived in a local variable that did not survive Streamlit's
  rerun. The project is now cached in st.session_state (keyed by the uploaded
  file's name/size), so the write step works; a new upload invalidates it.

v0.9.1

[0.9.1] - 2026-10-17

Added

• Milestone 7 (polish): pipeline.budgets (per-chunk and project character/
  duration/shot usage) and pipeline.lint (validator issues aggregated with
  project-level warnings). CLI gains --shots, --config, --strict, plus a
  budget/lint summary; the web wizard gains budget meters and a lint panel.
• README.md: project overview, status, install/usage, config, layout, plus a
  first-time setup section for install.sh/run_web.sh.
• run_web.sh: checks the venv/web extra is installed before launching the
  Streamlit wizard.
• install.sh: creates the project venv and interactively installs the chosen
  optional extra (none/pdf/web/both, also as an argument).

Changed

• pipeline.run clamps duration to the 5-10s range and shots_max to 1-3.

Fixed

• pyproject.toml: pin setuptools package discovery to script2prompt* so
  editable installs no longer fail on flat-layout dirs (out/, samples/,
  H3_Prompting_Docs/).

First pre-release version

[0.9.0] - 2026-10-16

Added

• PROJECT_PLAN.md: initial project plan (parsers, scene/chunk logic, character bible,
  multi-mode H3 support, web wizard, milestones).
• H3_Prompting_Docs/: local copies of the official Minimax H3 prompting guides
  (Text-and-Keyframe + Reference modes) used as the authoritative prompt spec.
• Scaffolding: script2prompt/ package (parsers/core/app stubs), pyproject.toml,
  config.toml, tests/ (stdlib unittest), samples/sample_scene.txt.
• script2prompt/core/config.py: stdlib tomllib loader with config.local.toml
  and OPENAI_BASE_URL / S2P_MODEL env overrides.
• script2prompt/core/llm.py: stdlib urllib chat-completions client (no openai SDK).
• Milestone 1 (offline core, T2VA): plain-text heuristic and Fountain-subset parsers
  (headings, cues with (O.S.) extensions, dialogue blocks, parentheticals,
  transitions); scene splitter with slugline context carry-over; character bible
  (speaker IDs in first-vocalization order, descriptions from NAME (traits)
  action intros, None IDs for non-vocal characters); chunker (duration estimate,
  greedy pack to target, <=10s, <=3 shots, MM:SS.mmm stamps, long-monologue
  split with <scenetrans> markers); deterministic T2VA template builder
  (core/fallback.py via core/h3_builder.py); T2VA validator (fields, shots,
  timestamps, <d> tags, soundscape/music budgets, 7000-char cap);
  core/modes.py instruction lines for all modes. 22 stdlib unittest tests,
  including end-to-end on samples/sample_scene.txt.
• T2VA template restyled to the real-example layout (Example_Text2Video_Prompt.txt):
  field labels on their own lines, standalone style sentence opening the body,
  dialogue as Name (S1) says <d>[Lang]words</d>. (no colon, no space after the
  language tag, period outside </d>). Conventions recorded in AGENTS.md and
  PROJECT_PLAN.md §2.1.
• Milestone 2 (keyframe modes): mode-specific [Shot 1] openings and landing
  sentences for I2VA (anchor <Picture 1> forward), FL2VA (single-shot path
  from Picture 1 to Picture 2) and L2VA (converge on <Picture 1> in the final
  shot); chunk() takes mode and forces single-shot chunks for FL2VA;
  validator checks instruction lines (without counting [Shot N] inside them),
  the blank line, and flags multi-shot FL2VA; modes.py gains section schemas.
• Milestone 4 (LLM wiring): core/prompts.py with per-mode system prompts for
  all five modes (system_prompt(), shared H3 rules + mode specifics);
  core/llm.py gains messages() (system prompt + structured chunk/bible brief,
  exact instruction line, R2V reference slots) and a working rewrite() with an
  injectable chat_fn; h3_builder.build(..., rewrite=...) uses the LLM when it
  returns non-empty text and falls back to the deterministic templates on error
  or empty output. 15 new tests (52 total), including a stdlib fake
  OpenAI-compatible server exercising chat round-trip and live rewrite.
• Milestone 5 (FDX + PDF parsers): parsers/fdx.py maps Final Draft
  <Paragraph Type=...> XML (via stdlib xml.etree) to elements, joining
  <Text> runs, reading only the script <Content> (title page ignored),
  skipping unknown types, and including dual-dialogue; parsers/pdf.py
  extracts text with lazily-imported pypdf and delegates to the plain-text
  parser (missing-extra error names .[pdf]); parsers/__init__.py adds
  parse_file(path) suffix dispatch. 8 new tests (60 total).
• Milestone 6 (wizard + export): core/pipeline.py end-to-end glue
  (parse_bytes, run, characters_markdown, prompts_markdown, manifest,
  reference_manifest, export_pack, write_pack); mode-aware reference
  slots (ref_manifest.picture_slots for I2VA/FL2VA/L2VA); app/web.py
  Streamlit wizard (upload → mode/settings → editable bible → generate/export;
  lazy streamlit import) and script2prompt/__main__.py CLI
  (python -m script2prompt SCRIPT, templates by default, --llm opt-in);
  samples/one_page.txt 1-page sample. 11 new tests (71 total) covering all
  five modes end-to-end plus export pack and CLI.

Fixed

• Audit fixes: character bible no longer absorbs one-off ALL-CAPS words (TV/DAY/OK);
  bare action mentions join only on second mention, cue/intro names always join.
  Over-long single sentences now hard-split inside monologue chunking.
  requires-python corrected to >=3.11 (tomllib); removed dead
  [tool.unittest] config; PROJECT_PLAN.md no longer mentions pdfplumber.
• Milestone 3 (R2V full-reference): core/ref_manifest.py proposes per-chunk
  Subject/Audio slots (Subject 1 environment, characters in bible order, one
  voice-timbre <Audio N> per speaker; pictures/videos empty placeholders);
  core/r2v.py builds the 6-section template with task-type prefix, valid
  retention markers, shot lists, and <Subject N> (Sx) speaker heads (no IDs in
  retention); validator checks section order, markers, task-type prefix, and
  true label definitions (line-start only). 7 new tests, 37 total.
• Audit fixes (R2V): numeric audio ordering (S10 no longer sorts before S2);
  R2V [Shot 1] opening gains its missing camera sentence.
• Deep-audit hardening pass (B1–B13, B15 in IMPROVEMENTS.md). Chunker keeps a
  character cue atomic with its dialogue and re-emits the cue before split
  monologue continuations, so speaker IDs are never lost ("unseen speaker");
  the monologue threshold is derived from the 10s cap. Voiceover detection now
  reads extension and delivery independently (V.O. + parenthetical works) and no
  longer treats O.S./O.C. as voiceover. Plain-text parser: stricter sluglines
  (rejects EXT.-led dialogue, accepts mixed-case - TIME, recognizes EST.),
  FADE IN/FADE OUT/SMASH CUT transitions, and cue lookahead so lone
  ALL-CAPS action beats are not characters. Character traits are captured
  mid-line (with a traits-like guard). CLI honors --duration 0; malformed
  config.toml raises a friendly ValueError; binary inputs are rejected;
  validator cross-checks speaker IDs against the bible and flags unseen
  speakers; h3_builder.build rejects scene dicts. 19 new regression tests
  (tests/test_hardening.py), 90 total.

Changed

• Dependency policy: project-local .venv/ only (git-ignored); zero required
  third-party deps for core (stdlib dataclasses, tomllib, urllib, unittest,
  xml.etree). Optionals are lazy-imported: pdf (pypdf, used by parsers/pdf.py
  only) and web (streamlit, used by app/web.py only). Dropped pydantic,
  pyyaml, openai SDK, pytest, pdfplumber, fastapi/uvicorn.
  See PROJECT_PLAN.md §3.2.
• Config format config.yaml → config.toml (stdlib tomllib).
  PROJECT_PLAN.md documents the dependency use policy (§3.2).