feat: local-LLM workflow, dedup UX, queue management, and autosave

[emirikol1]

Note from the submitter

I built this branch while learning Novel OS on my own install. I may have changed workflows the upstream author intended — in several places I wasn't fully sure what the original design was aiming for, so I optimized for what made sense to me as a daily user (local LLM, background jobs, autosave, queue visibility). Apologies if any of that diverges from your vision. Happy to revert, rename, or rework anything that doesn't fit the project's direction.

---

What this PR adds (high level)

1. LM Studio / Ollama as the primary LLM path (Claude CLI backend removed from this branch).
2. A global LLM concurrency queue with reorder, cancel, and a unified job popover.
3. Background jobs for slow AI work (dedup scans, chapter mining) so the UI stays usable.
4. Codex tooling in the dashboard (cast, plots, story bible) with dedup, generators, and backups.
5. UX polish: autosave, resume markers, AI-result stars, restart confirmation, modal portal fix.

All test fixtures use generic fiction names only. No user manuscript data, secrets, or home paths are committed.

---

Workflow markers

Marker	Where	Meaning
Blue dot ●	Binder / chapter board	One chapter only — last chapter you opened or ran an action in
Blue dot ●	Pipeline buttons (current chapter)	Last function you ran on this chapter (Revise, Mine plots, etc.)
Gold star ★	Tabs, buttons, binder rows	AI results ready to review (previews, dedup scans)

---

Intended workflow (how to use it)

Setup

• Point Novel OS at a local OpenAI-compatible server (LM Studio or Ollama).
• Set provider env in .env (gitignored).
• Start the app; open a project from Library.

Chapter pipeline

Outline → Draft → Revised → Validated → Approved → Final

Step	Button	What it does
Draft	Generate Draft	Scribe agent writes draft prose
Revise	Revise	Editor pass → Revised stage
Validate	Validate	Guardian continuity check
Approve	Approve	Marks chapter approved
Final	Promote → Final	Human-reviewed canonical text

Additional chapter tools: Regenerate, Expand placeholders, Outline from notes/text (preview → keep/discard), parallel Mine jobs for plots/characters/bible. Chapter text autosaves.

Dashboard codex

Tab	Features
Chapters	Board / Outliner; pipeline status dots; blue dot = last accessed chapter
Cast	Character editor with AI generate + autosave; Resolve duplicates
Plot Threads	Inline edit, drag-reorder, nest; Check subplot issues; autosave
Story Bible	Section editors + autosave; Deduplicate bible

LLM queue (sidebar → AI settings)

• Concurrent LLM requests — autosaved when changed.
• Queue popover — in-progress vs waiting, chapter badges, drag-reorder, cancel.
• Global system prefix — autosaved.
• Restart (flush queue) — confirms if background or LLM jobs are running before cancelling and restarting.

Backups & export

Explicit actions only: Backups modal, Export manuscript, Export project.

---

Feature list (for reviewers)

• Local LLM client; hardened JSON parsing for thinking-wrapped output
• FIFO LLM queue: concurrency limit, reorder, cancel, job labels
• Background job runner; non-blocking dedup and chapter mining
• Entity + bible dedup with disk-backed suggestions
• Plot panel issues; plot nest/reorder; title-row expand toggle
• Chapter regenerate / expand / outline (preview → keep/discard)
• Character & plot generators; project backup + portable export
• Modal portal fix; autosave across codex, chapters, and settings
• Workflow markers (binder vs button); gold AI stars; restart confirm
• Expanded pytest coverage

---

Security / privacy review (submitter)

• .env, *.db, outputs/, projects/, and logs are gitignored — not in this PR.
• Test fixtures use generic fiction names only.
• No API keys, home paths, or manuscript text in committed files.
• Single squashed commit; submitter author metadata anonymized.
• Triple-scanned before push: tree grep, full diff, commit metadata — all clean.

---

Test plan

• pytest tests/ from app root (137+ pass; 5 API job tests may need fresh process)
• LM Studio connected; Generate Draft on a test chapter
• Queue popover: in-progress vs waiting, reorder, cancel
• Blue dot: only one chapter in binder; button dot on last function used
• Gold ★ after background dedup or chapter preview
• Restart with jobs running → confirm dialog appears
• Autosave: character, plot thread, concurrency setting
• Project backup create + restore

[mrigankad]

Hi @emirikol1 — thank you for this, and I'm sorry it sat unanswered for five weeks. That's on me, not on the work.

I want to be straight with you rather than leave it drifting, so: I can't merge this branch as it stands, and I'd rather explain properly than let it sit open pretending otherwise.

Why not, honestly

1. It removes the claude_cli backend. I understand why — you were building around a local LM Studio setup and that path was in your way. But that backend is the reason the project exists in the shape it does: it runs the whole pipeline on an existing Claude subscription with no API key and no per-token billing, which is the one thing no comparable tool offers. It was the headline of v0.2.0. It has to stay.

To be clear, local models were never the problem — ollama, lmstudio and any OpenAI-compatible endpoint are all first-class in llm_client.py and always have been. Making them easier is welcome. Making them the only path isn't.

2. The branch is behind a rewrite. Since June the codebase has moved a long way — v0.3.0 landed a ProseMirror manuscript editor, track changes, a Codex, a compile engine, and a restructured UI. GitHub already reports this as CONFLICTING, and at +17,541 / −822 across 92 files the conflicts aren't the mechanical kind anyone can resolve confidently.

3. Size. Even without the drift, 17.5k lines spanning the LLM layer, job queue, dashboard and editor is more than I can review as one unit and honestly say I understood. That's a limitation of my review capacity, not a judgement on the code.

What I'd genuinely like to take

Several of your instincts were right, and some of them independently ended up in v0.3.0 — which is a decent signal you were reading the problem correctly:

• Background jobs for slow AI work — there's now a job registry; agents never run inside an HTTP request.
• Autosave — the chapter view debounce-saves and flushes on unmount.
• Codex tooling in the dashboard — cast, entries, relationships, portraits, and auto-extraction of entries from existing prose.

The ones still missing that your PR had and I'd welcome as focused PRs against current dev:

Idea	Why I want it
Job cancel + reorder	api/jobs.py can submit and poll but has no cancel. A long agent run you can't stop is a real problem.
A unified job popover	Job state is currently only visible on the page that started it.
Duplicate detection in the Codex	Auto-extract makes this matter more than it did in June, not less.
Local-LLM setup polish	Anything that makes LM Studio / Ollama easier to point at — as an addition to the provider list, not a replacement.

Any one of those as its own PR I can review properly and would likely merge. Small and separate beats comprehensive here.

On the note you opened with

You wrote that you might have changed workflows I intended, and offered to rework anything. That was a genuinely good way to open a large PR and I want to acknowledge it rather than let it pass — especially since I then left you waiting over a month for an answer.

I'm leaving this open for now in case you want to pull any single piece out of it. If you'd rather not, close it and I'll take no offence — and if you do open a smaller one, I'll review it in days, not weeks.