Thinkloom is a native, local-first writing studio that helps one writer move from conversation to ideas, draft, revision, and a versioned release while preserving an inspectable creative-process record.

The application is built with Tauri 2, React, TypeScript, Rust, SQLite, canonical Markdown/JSON/JSONL files, hidden Git checkpoints, operating-system credentials, and atomic exports. There is no hosted or browser companion build.

OpenAI Build Week

Thinkloom was created from scratch on July 16, 2026, during the OpenAI Build Week submission period. Christopher Chambers is the individual entrant and project owner. The project was planned and implemented with Codex using GPT-5.6 Sol at High reasoning effort.

How Christopher and Codex collaborated

Christopher set the product direction and made the key product decisions: serve individual writers; make the experience conversation-first; keep model output preview-first and user-controlled; preserve creative provenance without reducing authorship to a percentage; use a native Tauri application; retain no audio; and remove the unintended hosted companion so the product remained native-only.

Codex accelerated the work by:

translating the MVP documents into a clickable workspace framework;
turning the approved interface into a gated implementation plan;
implementing the React/TypeScript interface and Rust/Tauri service layer;
proposing and integrating TipTap for structured editing and Markdown round-tripping;
implementing SQLite state, canonical files, atomic writes, provenance hashing, hidden checkpoints, provider boundaries, exports, backups, and ZIP validation;
responding to Christopher's design and product-boundary feedback; and
running type, lint, unit, native, packaging, and launch verification.

GPT-5.6 Sol was used throughout the Codex discussions for product reasoning, architecture, implementation, debugging, and verification. The dated Git history and the primary Codex build task document work performed during the submission period.

Primary Codex build task: 019f6edd-de3d-78d2-8b6f-67af5fe6c44c.

Required /feedback session ID: 019f6edd-de3d-78d2-8b6f-67af5fe6c44c.

See CODEX_COLLABORATION_LOG.md for the complete project conversation record. See HACKATHON_SUBMISSION.md, DEMO_SCRIPT.md, and JUDGING.md for the submission package.

Implemented workflow

Reversible Ideation, Drafting, and Finalization phases
Three-pane Phase 1 workspace for session context, live conversation, and idea distillation
Persona, genre, lore, and provider-aware web-search guidance assembled into each conversational instruction
Per-message Append to Ideas controls, editable drafting paper, one-off LLM distillation, and structured session export
Typed conversation, challenge levels, push-to-talk transcription, persona-linked speech output, and no retained audio
Suggested ideas with explicit accept/reject, editing, variants, archiving, source links, drafting sets, and merges
TipTap/ProseMirror structured manuscript editor with canonical Markdown round-tripping, undo/redo, headings, lists, selection replacement, and cursor insertion
Persisted preview-first generation states with retry-safe provider errors and partial acceptance
Ollama, OpenAI, and OpenAI-compatible provider profiles; credentials use the operating-system vault
Local/Cloud/Mixed status and project-scoped cloud approval
SQLite live state, atomic canonical files, seven rotating recovery snapshots, and a hidden Git repository per project
Append-only SHA-256 provenance journal with chain-head verification and a contribution relationship view
Named versions, restore controls, release checkpoints, and tags using non-Git language in the UI
Markdown, HTML, PDF, plain text, sanitized evidence ZIP, and complete project backup ZIP generation
ZIP import path, symlink, file-count, and expanded-size validation
Responsive, keyboard-navigable, screen-reader-labeled UI with reduced-motion and dark-mode support

Development

Requirements: Node.js 22.13 or newer, Rust 1.77.2 or newer, Git, and the Windows WebView2 runtime for the primary desktop target.

npm install
npm run dev
npm run build
npm run tauri -- dev
npm run tauri -- build

Quality checks:

npm run typecheck
npm run lint
npm test
cd src-tauri
cargo fmt --check
cargo test

Judge testing

Judges should not need to rebuild Thinkloom. A public Windows x64 installer release will be linked from JUDGING.md before submission. No Thinkloom account is required. The included sample project demonstrates the primary workflow without credentials; live model requests additionally require a local Ollama instance or an entrant-supplied OpenAI/OpenAI-compatible credential.

Project storage

A Thinkloom project is self-contained. Canonical files live under manuscript/, ideas/, conversations/, provenance/, and style/. Live SQLite state and rotating snapshots are under .thinkloom/ and are excluded from the project's hidden Git history. Audio retention is always false; no audio file extension is created by the native service.

Prompt configuration

Every language-model prompt is editable JSON. Thinkloom creates documented user copies of conversation.json and drafting.json and exposes their folder under Settings → Prompt configuration. The files reload before every request, so technical users can iterate without rebuilding or restarting. Invalid JSON or missing template variables stop the request with a recoverable error and leave the writing unchanged.

See PROMPTS.md for each file's effect, variables, editing workflow, validation behavior, and reset procedure.

Provenance architecture

The approved Thinkloom 1.0 provenance architecture is documented in the Stage 1 normative provenance specification. It defines the future single-writer subsystem, canonical records, segmented ledger, recovery protocol, retention policy, native verification, and release bindings.

Thinkloom 0.5.0 extends the formal Stage 2 package with canonical provenance assertions, point-in-time evaluations, versioned semantic registries, and deterministic invalidation vectors. The current native provenance writer must not be represented as conforming until the later implementation and fault-injection stages are complete.

Provider setup

Ollama defaults to http://127.0.0.1:11434 and model llama3.2. OpenAI and compatible credentials are entered in Settings and saved through Windows Credential Manager, macOS Keychain, or Linux Secret Service. The first cloud operation in each project requires explicit approval.

See IMPLEMENTATION_STATUS.md for release gates that require external models, signing credentials, or additional platform validation.

README.md

Thinkloom