physics-IDE

physics-IDE is a Linux-first desktop research environment for developing, comparing, and testing scientific theories in a structured workspace. It combines a Tauri-based desktop shell with a lightweight web frontend and Rust-backed analysis utilities.

Project Goals

The immediate direction for physics-IDE is to focus on a Linux-exclusive build and Debian-based deployment workflow. Windows support is currently deferred while the compilation process remains too complex to maintain effectively.

Version 8 is the release-candidate baseline. The next major milestone is version 9, the first official physics-IDE release, focused on production readiness and efficient project-aware AI operation.

The version 8 vision is to provide a flexible environment for:

• organizing theory material, notes, equations, and manuscripts in one place;
• ingesting theory sources from different paradigms without rejecting them up front;
• generating a structured master manuscript from imported markdown theory content;
• building a project knowledge index so the AI can navigate chapters, sections, and subsections coherently;
• linking theory content to reusable tools, scripts, experiments, and datasets;
• supporting educational and exploratory workflows across mainstream, hybrid, and non-standard theory families;
• connecting theory content to empirical data and transparent evaluation workflows;
• shipping a polished Linux desktop experience with reliable .deb packaging and installation.

Current Progress

The project has moved from a simple desktop shell toward a more practical theory-development workspace.

Latest progress (2026-10-17)

• sandboxed Linux RUN / CALC execution and added safeguards against unsupported AI claims that files were written.

Latest progress (2026-10-15, v9.0.2)

• cloud chat (OpenAI, Gemini) no longer aborts replies after 20 s: 20 s connect limit, 15 min per reply, and request errors name the cause (timeout / connect / TLS) — long Task rounds on gpt-4.1 failed with a bare "error sending request";
• Settings ▸ Theory Source File is now saved with the other paths.

Latest progress (2026-10-15, v9.0.1)

• Task mode DONE gate: <<<DONE>>> is rejected (max two send-backs) while files written during the task hold TODO/placeholder code, a report cites RESULT: lines no run printed, or the requested ±err is missing;
• Local AI replies stream with no total time cap (Settings ▸ Local AI silence limit, default 60 min), live pane progress, Stop cancels, distinct timeout / not-running / crash errors;
• View ▸ Monitors: CPU, Memory and System floating windows (incl. the Ollama process).

Latest progress (2026-10-15)

• cut the v9.0.0 release line: tag-driven .github/workflows/release.yml builds the Linux .deb with the Tauri bundler and publishes it (plus a stable physics-ide_latest_amd64.deb) to GitHub Releases, linked from gpd-research.com;
• completed the interactive smoke test of the agentic subsystems and shipped the fix batch: hypothesis-preamble leak, append-only project_memory.md, correct token-budget stop reason, clean task stop on provider errors, "Settings" wording everywhere;
• moved every remaining disk / network / subprocess Tauri command off the UI thread and added a Math Fixer progress bar;
• added a Word Processor picker (detected apps + custom command) in Settings ▸ System & Apps used by "Open in word processor";
• retrieval hygiene: app-generated reports are no longer indexed and the mechanism-intent boost requires real topical overlap;
• Manuscript Tools open full-screen and gained a Master Axiom mode with custom source, output directory and md / pdf / docx format;
• Local AI: quantization selector (q4_0, q4_K_M, q5_K_M, q8_0, fp16) with per-quant size and fit estimates;
• earlier in this cycle: READ / APPEND / EDIT / allowlisted RUN file protocol with guardrails, SEARCH and REMEMBER, shadow-git checkpoints with revert, Task mode with budgets and Pause / Approve / Stop, CALC scratch computation with inline figures, per-pane Thread tools.

Latest progress (2026-10-11)

• added Local AI (Ollama) as a third per-pane provider alongside Gemini and OpenAI, configured through the same AI Configuration dialog, with non-streaming /api/chat requests, 16K context, and local token-usage reporting;
• added a "Check Hardware Specs" probe (CPU, RAM, VRAM via nvidia-smi/rocm-smi, free disk on the Ollama model volume) and a curated model catalog that estimates per-query response time against a ~9K-token primer prompt, shows a quick↔deep spectrum marker, and marks models "Unusable" when they exceed 60 s, do not fit memory, or would fill the disk to 95%;
• added one-click model installation that launches the configured terminal with the Ollama install/pull commands, and shows downloaded models in green with Left/Right pane assignment;
• added macOS support to the Local AI subsystem: sysctl/vm_stat hardware probing, Apple Silicon unified-memory estimates by chip tier, Homebrew-based install, and open -a Terminal launching, so a Mac build can run local models;
• added a unified "Document ▾" actions menu (note pad, built-in editor, session notes, session equations, app buffer, external text editor or word processor) on the file tree, chat bubbles, recall viewer, note pad, equation panel, and document modals, plus a glass-styled built-in editor and a "Session Docs" top menu;
• added PDF text extraction for the note pad and editor, and an in-app document buffer so equations survive copy/paste without clipboard mangling;
• added SQLite thread autosave with a per-pane "Recall session" dropdown that opens the three most recent saves in a read-only viewer (no tokens spent) with Close / Copy to session notes / Load into thread;
• added a manual ⟳ file-tree refresh (no background watcher), code-fence rendering with Copy buttons, per-message Copy and "Add to notes" (session_notes_YYYY-MM-DD.md in the project root), and a token-limit freeze banner so a thread is preserved instead of lost;
• added a Math Fixer mode to Manuscript Tools that scans a file, directory, or the whole project for Unicode math symbols, spelled-out Greek letters (with upper/lower-case choice), bare identifiers outside math, Python/shorthand notation, and malformed delimiters, with per-finding review, .bak backups, and dry run; PDF export now prefers xelatex → lualatex → pdflatex;
• rewrote master-axiom generation as a deterministic LaTeX compilation of every governing equation with labels and variable definitions (no provider rewrite, no placeholder prose), injected whole into the startup primer, with the manuscript itself delivered by bounded per-message retrieval;
• attached the budgeted local retrieval packet to every chat message and weighted file names and headings in ranking so a directly named appendix wins;
• gave the AI a <<<FILE path="...">>> protocol for creating and writing files inside the project root when read/write access is enabled, with OK/FAILED reports fed back to the model.

Latest progress (2026-08-16)

• completed v9 prompt assembly work with deterministic stable-to-dynamic ordering, preserved provider roles, source provenance, per-thread token estimates, and provider-reported usage diagnostics;
• validated native OpenAI prompt caching on gpt-4.1, gpt-4.1-mini, and gpt-4o-2024-08-06, each reporting 1,024 cached tokens and an 81% cache-hit ratio for the tested shared prefix;
• added the AI Request Inspector for prompt ordering, stable-prefix fingerprints, source contributions, cache probes, and provider usage totals;
• replaced recursive briefing-time workspace scans with compact visible-tree export and root-only fallback behavior;
• added deterministic physics-ide.structural-context/v1 compilation with stable IDs, typed scientific records, source provenance, validation, and compact-core benchmarking;
• used blind A/B testing to reject full-corpus structural-core replacement for the BMI project when it was both larger and less useful than concise legacy context;
• added a project-keyed local SQLite FTS5 retrieval index under Tauri application data, with incremental refresh, bounded Unicode-safe chunks, explanatory neighbors, generated-context exclusions, safe local deletion, and legacy evidence fallback;
• validated the local vector foundation with statically linked sqlite-vec, a 384-dimensional cosine index, and FastEmbed model metadata without network-dependent tests;
• added an explicit 91 MB local model installer with immutable revision pinning, per-file SHA-256 verification, app-local storage, and a required offline inference activation probe;
• connected index refresh to incremental local embeddings with stable-ID reuse, transactional changed/deleted-vector invalidation, and inspector diagnostics;
• added hybrid retrieval using symmetric reciprocal-rank fusion across FTS5 and cosine-vector candidates, cached local inference, rank provenance, and lexical fallback;
• added deterministic typed graph-neighbor expansion from explicit project-source relation statements, with bounded provider context and inspector provenance;
• added an eight-case, provider-free four-family retrieval benchmark comparing Recall@3 and MRR@3 with acceptance thresholds, graph coverage, and per-case provenance;
• enforced configurable backend retrieval budgets with complete inclusion, exclusion, truncation, freshness, model, and source diagnostics;
• stopped automatic primer transport of master-axiom/project-awareness corpus excerpts while retaining explicit legacy recovery and comparison modes;
• added complete local-index lifecycle controls, non-mutating integrity/freshness inspection, versioned schema validation, and corruption/incompatibility quarantine recovery;
• replaced generated-primer editing/raw sync with a read-only Awareness Inspector, canonical refresh, and explicit bounded legacy recovery;
• restored idea-pad sync as an independent dynamic session-notes slot with visible-tree provenance while retaining Save/Save As recovery for user notes;
• moved model installation, indexing, rebuild, query, and benchmarks to blocking workers so large local projects remain responsive;
• improved natural-language mechanism retrieval with generic derivation terms, duplicate-copy collapse, direct-source preference, and heading-level multi-hit manuscript search;
• fixed a fatal UTF-8 manuscript-search crash caused by byte slicing across multibyte punctuation and added result content previews;
• added theory-neutral governing-mechanism reranking and Previous/Next section navigation for manuscript search results;
• fixed request-inspector/provider serialization failures from lone UTF-16 surrogates while preserving valid emoji and code-point-safe tree truncation;
• restored refreshed awareness to the reusable prompt prefix and added observable idea-pad session-note synchronization;
• expanded the backend regression suite to 74 passing tests plus an explicit model-install, real-vector persistence, hybrid-query, and real-model benchmark probe.

Latest progress (2026-08-11)

• added an AI file-access permission layer that limits file creation and modification to the active project root unless the user explicitly changes that root in Customize;
• made the AI advisory-only by default in the briefing packet language, so the app clearly distinguishes project reasoning from direct local workspace editing;
• replaced the decorative workspace tree output with a compact AI-friendly markdown project map that is easier for the AI to parse and reason over;
• added a read-only / read-write toggle in the settings panel so project-aware AI operations can be safely gated by user intent;
• upgraded chat input to support Shift+Enter line breaks and a more natural multiline workflow for structured prompting;
• added file attachment support for prompt threads, opened by default from the active project root and including text/image payload content in the prompt sent to the AI;
• tightened the user-facing idea-pad workflow so a session note can sync directly into AI context with an optional visible project-tree scope filter;
• added the Oceanic theme and refined Chromostereopsis for longer, more comfortable dark-theme use;
• verified the backend remains stable with the current regression suite passing after the AI access-control, prompt-attachment, and UI polish changes.

Latest progress (2026-08-10)

• completed Gemini and OpenAI route validation hardening in Advanced AI Routing so save operations are blocked unless both pane routes validate;
• migrated Gemini defaults and route handling away from retired 2.0 IDs toward current-generation model selection behavior;
• added provider model catalog revision with validation so users can auto-filter to models that actually pass with their active key/project;
• added live per-provider validation progress feedback (countdown + progress bar) so long model sweeps visibly advance instead of appearing frozen;
• refactored Advanced AI Routing into a two-column layout (catalog left, pane controls right) for improved laptop usability;
• reduced translucency opacity across the glass UI surfaces for a lighter, more modern Linux desktop feel;
• completed regression coverage for routing UI/layout and catalog validation controls, with current tests green.

Latest progress (2026-07-30)

• added a built-in Markdown Documents viewer with rendered preview, fuzzy search, and editor launch workflow;
• added single-document PDF export from Markdown Documents with output-directory selection;
• connected Markdown Documents and Manuscript Tools with cross-navigation buttons for rapid workflow switching;
• improved manuscript rendering/export behavior so PDF and DOCX are generated through Pandoc conversion;
• added explicit GitHub Username/PAT settings and aligned markdown save behavior with configured GitHub mode vs local-save mode;
• expanded the in-app Help system with:
  • GUI Button Glossary,
  • Push/Pull Context and Common Errors,
  • Startup Initial Setup Workflow and Checklist;
• completed a UI housekeeping pass with terminology alignment and broad tooltip coverage (including keyboard Enter hints on chat/search flows).

Today's progress (2026-07-28)

• confirmed reliable Gemini model communication from the desktop app;
• fixed markdown file-opening from the project tree view;
• established the version 7 direction around project-aware AI memory and theory indexing;
• prepared the groundwork for a new Tools menu and manuscript-ordering workflow.

Implemented so far

• a Tauri desktop app shell with a configurable interface and integrated terminal;
• settings for project roots, theory directories, master-axiom paths, and AI/provider configuration;
• a master-axiom generation flow that scans theory markdown content and produces a structured draft;
• a theory import pipeline that can ingest a source file and split plain-text manuscripts into markdown sections;
• an initial theory-mode classification layer that recognizes mainstream, hybrid, and left-field-style content;
• regression tests for scientific template generation, theory-style classification, and manuscript import.

Current focus

• verify the first .deb from the release workflow installs and launches cleanly; then iterate on user feedback from the v9 release;
• make local Ollama models the normal operating mode (one local pane at a time on 16 GB machines), with cloud providers as the high-capacity option;
• "Agentic Build-out" documentation, then watch-mode experiment sweeps and a self-verification pass;
• add a Table Fixer and an HTML/KaTeX PDF export route for table-heavy manuscripts;
• implement lazy directory token-cost guidance and complete the production release audit.

Version 8 Implementation Order

1. Project knowledge index

   • Parse the master manuscript and theory markdown directory into a structured topic tree.
   • Generate chapter, section, and subsection summaries that can be used by the AI as a compact navigation layer.

2. Compact project digest

   • Create a token-efficient digest file that summarizes the theory corpus, assumptions, tools, and experiments.
   • Use this digest as a prompt context layer for AI sessions.

3. Manuscript composition workflow

   • Allow the user to reorder imported markdown files into a preferred master-document sequence.
   • Support logical sorting for numbered sections and appendix-style files.
   • Render a combined markdown document from the chosen order.
   • Export the result as Markdown, PDF, or DOCX from a new Tools menu workflow.
   • Support an optional AI training export that writes a replacement training artifact for project-aware AI context.

4. Tool and experiment registry

   • Add a project-level registry of reusable scripts, notebooks, and analysis tools.
   • Track prior experiments and link them to the theory topics they support.

5. AI awareness integration

   • Inject the project digest, topic index, and tool registry into the AI briefing pipeline.
   • Make the AI prefer existing tools and previous analyses before proposing new ones.
   • Support in-thread file attachment so users can provide a selected document or image directly to an AI lane.

6. UI polish and workflow consolidation

   • Add the new Tools menu and move version-7 functions into that drop-down as they are introduced.
   • Keep the Customize menu focused on configuration paths and app settings.
   • Expand the terminal area in the left wing to make the workspace tools more usable.

Development

Prerequisites

• Node.js and npm
• Rust toolchain
• Tauri prerequisites for your OS
• Pandoc (required for real PDF and DOCX manuscript/markdown export)

Run locally

npm install
npm run tauri dev

Verify backend tests

cargo test --manifest-path src-tauri/Cargo.toml

Notes

This repository is under active development. The current implementation is intentionally modular so new theory parsers, empirical evaluators, and scientific workflows can be added over time.

Release Tracking

• Desktop stabilization toward v6.0.0 is tracked in docs/releases/v6-desktop-checklist.md.
• UI-to-function mapping coverage is tracked in docs/releases/v6-control-map.md.
• The v8 release-candidate plan is tracked in docs/releases/v8-model-flexibility-plan.md.
• The v9 first-official-release goals and execution roadmap are tracked in docs/releases/v9-official-release-goals.md.

API Key Transparency

• Provider API keys entered in settings are stored locally on the device in encrypted form.
• Keys are decrypted only when needed for provider requests and are not displayed back in plain text in the UI.
• Keys are used only to send requests to the provider selected in the UI.

Version 8 Release Candidate

Version 8 marks the current release-candidate milestone for physics-IDE.

• Project-aware AI behavior is stable enough to materially outperform a generic side-by-side browser LLM workflow for in-project theory work.
• The desktop workflow now supports a coherent paradigm for theoretical-physics modeling, iteration, and testing, with AI carrying repetitive context-heavy tasks while the human remains the primary director of theory evolution.
• The app is now tuned for a clean AI-first project workflow: idea-pad driven prompting, scoped project context, safe local file access guardrails, and a more natural desktop UX.

Version 8 Goals

Version 8 is focused on refinement, flexibility, and polish.

• model freedom: user-selected OpenAI and Gemini model IDs without code edits;
• workflow consolidation: migrate mature workflows into the top-menu Tools dropdown to free left/right wing real estate;
• layout control: expand View controls so users can toggle pane elements such as file tree and primer-related surfaces;
• primer simplification: evaluate how much primer work can be automated by project-aware context, including an idea-pad-driven pathway that can append daily notes into primer context;
• UX coherence: keep customization centered on path/location setup while reducing repetitive manual context assembly.

Tracking references:

• docs/releases/v7-release-checklist.md
• docs/releases/v7-release-notes.md
• docs/releases/v8-model-flexibility-plan.md
• docs/releases/v9-official-release-goals.md

Ubuntu Linux Full Build Guide (v8 RC)

Use this sequence on a local Ubuntu laptop for a clean production build.

1. Install system dependencies

sudo apt update
sudo apt install -y \
   build-essential \
   curl \
   wget \
   file \
   pandoc \
   libgtk-3-dev \
   libayatana-appindicator3-dev \
   librsvg2-dev \
   patchelf \
   libwebkit2gtk-4.1-dev

2. Install Node.js 20 LTS (if not already installed)

curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash -
sudo apt install -y nodejs
node -v
npm -v

3. Install Rust toolchain (if not already installed)

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
source "$HOME/.cargo/env"
rustc -V
cargo -V

4. Clone and install project dependencies

git clone https://github.com/GPD-Research/physics-ide.git
cd physics-ide
npm install

5. Run automated checks before packaging

node --test src/ai-config.test.js
cargo test --manifest-path src-tauri/Cargo.toml

6. Build production desktop artifacts

npm run tauri -- build

7. Locate artifacts

• Debian package and related artifacts are produced under:
  • src-tauri/target/release/bundle/

8. Install local Debian package (if generated)

sudo dpkg -i src-tauri/target/release/bundle/deb/*.deb
sudo apt -f install -y

9. Launch and smoke-check

• open the installed app;
• verify workspace loading, AI provider settings, and AI Testing modal flows;
• run one context probe and confirm report generation/open-report behavior.

If build issues appear, capture full logs:

npm run tauri -- build > build.log 2>&1
tail -n 120 build.log