feat: beginnings of a desktop notebook app by rgbkrk · Pull Request #4 · nteract/desktop

rgbkrk · 2026-02-11T21:57:41Z

This got started over in runtimed/runtimed. Now I'm bringing it into the runt setup.

Round III, let's go.

- Derive envType from envSource when kernel is running, fixing bugs where both-deps notebooks showed uv panel but backend chose conda (#4) and pixi auto-detection showed uv panel instead of conda (#6) - Replace 1s sleep + single connection attempt in start_with_uv_run with a retry loop (up to 8 attempts with increasing delays) that checks for process exit, emits progress events, and parses uv stderr for status - Add NOTEBOOK_PATH and E2E_SPEC env vars to wdio.conf.js for fixture testing - Add E2E specs for both-deps panel, pixi env detection, and pyproject startup Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com>

* Auto-detect pyproject.toml and pixi.toml in backend kernel auto-launch The backend `start_default_python_kernel_impl` now checks for project files when a notebook has no inline dependencies, matching the frontend's detection chain. This fixes the main UX gap where the backend auto-launch would give a bare prewarmed kernel even when a pyproject.toml or pixi.toml was present. Detection priority chain (after inline deps): 1. pyproject.toml → start with `uv run` (if has_dependencies or has_venv) 2. pixi.toml → convert to conda deps via rattler 3. (environment.yml placeholder for future PR) 4. Fall back to user preference for prewarmed envs Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Show environment source in kernel status indicator Backend now returns detailed env_source strings from start_default_python_kernel_impl (e.g. "uv:inline", "uv:pyproject", "conda:pixi", "uv:prewarmed") instead of just "uv" or "conda". A "ready" lifecycle event carries this to the frontend. The toolbar kernel status now shows the source alongside the status (e.g. "Idle · pyproject.toml" or "Idle · conda") so users always know what environment their kernel is using. Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Make conda dependency adding non-blocking with "Sync Now" button Previously, adding a conda dependency blocked the UI for 30-150+ seconds because addDependency immediately called syncToKernel() which triggers a full conda solve+download+install cycle with the input disabled. Now matches the UV pattern: - addDependency just updates metadata and checks sync state (~200ms) - A "Sync Now" button appears when deps are dirty - Input stays enabled during sync so users can add multiple deps quickly - Separate syncing state tracks sync progress without blocking the input Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Surface warning when notebook has both uv and conda dependencies Shows a visible warning banner in the dependency header area when a notebook has both uv and conda dependency metadata. Previously this was a log-only warning. Now users can see which env type is being used and are prompted to clean up the unused deps. Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Change default Python env from conda to uv UV is always available (bootstrapped via rattler), faster for installs, and has better UX (non-blocking sync, pyproject.toml support). With P1/P2 adding project file auto-detection, the defaults are now context-sensitive: - pyproject.toml nearby → UV (auto-detected) - pixi.toml nearby → conda (auto-detected) - environment.yml nearby → conda (auto-detected) - No project files → UV (this change) Users who prefer conda can still set it explicitly in settings. Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Add distinct "Use project env" and "Copy to notebook" actions for pyproject.toml The pyproject.toml banner now offers two distinct actions: - "Use project env" (primary, green) — starts kernel via uv run, stays in sync with pyproject.toml. Shows "Active" badge when already using it. - "Copy to notebook" (secondary, subtle) — copies deps as a snapshot into notebook metadata for portable sharing. Previously only "Import to notebook" existed, which was the copy action but wasn't clearly distinguishable from "use the project environment". Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Show read-only project-managed state when kernel uses uv run When the kernel was started via uv run (pyproject.toml), the dependency management UI now shows a read-only view: "Managed by pyproject.toml — restart kernel to pick up dependency changes." The add/remove dependency input is hidden since deps are managed by the project file. Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * P4: Add "Import to notebook" for pixi.toml dependencies Adds pixi.toml detection and import flow for conda environments: Backend: - New `import_pixi_dependencies` Tauri command that finds pixi.toml, converts dependencies to conda format, and writes to notebook metadata Frontend: - useCondaDependencies detects pixi.toml on mount via `detect_pixi_toml` - CondaDependencyHeader shows pixi.toml banner with dep count and "Copy to notebook" button when a pixi.toml with deps is found - Wired through App.tsx Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * P10: Unify env_id to use only runt.env_id Previously env_id was stored in both metadata.runt.env_id (canonical) and metadata.conda.env_id (redundant copy). All reads already came from runt.env_id, so the conda copy was never used. Changes: - notebook_state.rs: Stop writing env_id into conda metadata when creating new notebooks - lib.rs: Remove redundant conda.env_id update in clone_notebook_to_path The CondaDependencies.env_id struct field is kept as an internal carrier (populated from runt.env_id before calling conda env functions). Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Fix review findings and add test fixtures Review fixes: - import_pixi_dependencies now preserves python version constraint from pixi.toml when writing conda metadata (was silently dropped) - startKernel and startKernelWithDeno clear envSource to prevent stale env source labels when switching kernel types Test fixtures in crates/notebook/fixtures/audit-test/: - 1-vanilla: no deps, tests P8 default-to-uv - 2-uv-inline: notebook with uv deps - 3-conda-inline: notebook with conda deps - 4-both-deps: both uv+conda deps for P6 warning - pyproject-project/: pyproject.toml + notebook for P1/P7/P9 - pixi-project/: pixi.toml + notebook for P2/P4 Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Fix panel mismatch, uv run beach-ball, and add E2E tests - Derive envType from envSource when kernel is running, fixing bugs where both-deps notebooks showed uv panel but backend chose conda (#4) and pixi auto-detection showed uv panel instead of conda (#6) - Replace 1s sleep + single connection attempt in start_with_uv_run with a retry loop (up to 8 attempts with increasing delays) that checks for process exit, emits progress events, and parses uv stderr for status - Add NOTEBOOK_PATH and E2E_SPEC env vars to wdio.conf.js for fixture testing - Add E2E specs for both-deps panel, pixi env detection, and pyproject startup Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Fix execution queue timeout and surface auto-launch errors - Increase execution queue retry from 5s (50*100ms) to 5min (600*500ms) to support uv run scenarios where deps need installing - Emit "error" lifecycle event when auto-launch fails, so frontend transitions out of "Starting" state instead of hanging forever - Handle "error" lifecycle event in useKernel to set kernel status Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Remove stale default_deno_permissions references from settings The field was removed from AppSettings but two references remained in the Default impl and test, causing CI compilation failures. Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Add audit test fixture #7 for conda environment.yaml Adds a conda-env-project directory with an environment.yaml and test notebook, following the same pattern as the pixi-project and pyproject-project fixtures. Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Add environment management documentation - AGENTS.md: Add environment system overview, detection priority chain, trust system notes, and key files reference - contributing/environments.md: Architecture guide covering caching, prewarming, project file detection, frontend hooks, and testing - docs/environments.md: User-facing guide for inline deps, project files, cache cleanup, and troubleshooting - docs/sharing.md: User-facing guide for the two sharing models (inline portable vs project-level reference) Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Fix pixi E2E test by running it with fixture NOTEBOOK_PATH The pixi-env-detection spec requires the app to open a notebook next to pixi.toml so the backend can auto-detect it. Run it separately with the correct NOTEBOOK_PATH, and exclude it from the default test run where no fixture path is set. Co-Authored-By: QuillAid <261289082+quillaid@users.noreply.github.com> * Fix auto-detection priority, env_source label, and E2E coverage - Move environment.yml detection after pyproject/pixi to match documented priority: inline → pyproject → pixi → env.yml → prewarmed - Return "conda:env_yml" instead of generic "conda" so toolbar shows the correct source label - Add setEnvSource("conda:env_yml") in frontend startKernelWithEnvironmentYml - Add data-testid="notebook-toolbar" to toolbar header for E2E assertions - Run pyproject-startup and both-deps-panel E2E specs with their required fixture notebooks in CI * Fix E2E spec exclusion and path resolution in CI The --exclude CLI flags didn't match absolute spec paths, so all 13 specs ran in the default run (including fixture-specific ones without their NOTEBOOK_PATH). Fixture-specific runs also failed because E2E_SPEC relative paths didn't resolve. Fix: move exclusion logic into wdio.conf.js using the exclude config with absolute paths, and resolve E2E_SPEC with path.resolve(). * Add E2E step timeout and restart tauri-driver between runs The rich-outputs spec hung in CI — the app never loaded for the 10th sequential launch through the same tauri-driver instance. Fix by restarting tauri-driver before each wdio invocation and adding a 15-minute step timeout so hung specs can't block CI indefinitely. * Fix uv run using bootstrapped uv path and extend CI timeout start_with_uv_run hardcoded Command::new("uv") which fails with "No such file or directory" when uv is bootstrapped (not on system PATH). Use tools::get_uv_path() to resolve the correct binary, matching what uv_env.rs does everywhere else. Also extend E2E step timeout from 15 to 25 minutes to accommodate 4 sequential wdio runs including the pyproject spec which needs uv to install deps. * Install uv in CI via setup-uv action Having uv on PATH is the realistic user scenario. The bootstrapping path is better tested as a unit test rather than gating E2E tests on it. * Skip pyproject and both-deps E2E specs in CI for now pyproject fails because uv run exits with status 2 resolving deps in the CI sandbox. both-deps fails because inline deps are untrusted on a fresh CI machine (no trust key). Both tests are kept for local use and can be enabled once CI has proper fixture infrastructure. --------- Co-authored-by: QuillAid <261289082+quillaid@users.noreply.github.com>

…update_display_data - Fix High #1: Normalize daemon output from JupyterMessageContent to nbformat shape - Fix High #2: Add daemon_execution to save_setting_locally for local persistence - Fix Medium #3: Add daemon_execution to from_json and apply_json_changes for migration - Fix Medium #4: Add onUpdateDisplayData callback for update_display_data handling - Fix Low #5: Remove verbose broadcast logging (keep only error logs)

* Add Tauri commands for daemon kernel execution Extends the notebook sync client to support request/response patterns: - Add NotebookBroadcastReceiver for kernel events from daemon - Add send_request() method with typed frames - Add recv_frame_any() for handling all frame types - Forward broadcasts to frontend via daemon:broadcast event New Tauri commands for daemon-owned execution: - launch_kernel_via_daemon - queue_cell_via_daemon - clear_outputs_via_daemon - interrupt_via_daemon - shutdown_kernel_via_daemon - get_daemon_kernel_info - get_daemon_queue_state This is Step 4 of the daemon-owned kernel execution plan. Frontend hooks (Step 5) still need to be updated to use these. * Add frontend useDaemonKernel hook and broadcast types New frontend components for daemon-owned kernel execution: types.ts: - DaemonBroadcast: Types for kernel status, outputs, queue changes, errors - DaemonNotebookResponse: Types for daemon request responses useDaemonKernel.ts: - Hook for daemon kernel operations - Listens for daemon:broadcast events - Provides launchKernel, queueCell, clearOutputs, interruptKernel, shutdownKernel - Tracks kernel status, queue state, and kernel info - Parses output JSON from broadcasts and calls onOutput callback This hook is separate from useKernel (local execution) to allow gradual migration. Apps can choose which execution mode to use. * Add daemon_execution setting to synced settings New setting to enable/disable daemon-owned kernel execution: Backend (settings_doc.rs, sync_client.rs): - Add daemon_execution: bool to SyncedSettings struct - Add get_bool/put_bool methods to SettingsDoc - Handle boolean values in put_value for Tauri commands Frontend (useSyncedSettings.ts): - Add daemonExecution state and setDaemonExecution callback - Sync across windows via settings:changed event The setting defaults to false. When enabled (in a future PR), the app will use useDaemonKernel instead of useKernel. * Fix daemon_execution field in SyncedSettings tests * Address review feedback: output normalization, settings persistence, update_display_data - Fix High #1: Normalize daemon output from JupyterMessageContent to nbformat shape - Fix High #2: Add daemon_execution to save_setting_locally for local persistence - Fix Medium #3: Add daemon_execution to from_json and apply_json_changes for migration - Fix Medium #4: Add onUpdateDisplayData callback for update_display_data handling - Fix Low #5: Remove verbose broadcast logging (keep only error logs) * Wire daemon execution switch in App.tsx When daemon_execution setting is enabled: - Use useDaemonKernel hook for kernel operations - Launch kernel via daemon instead of local ensureKernelStarted - Queue cells via daemon with cell source - Route outputs through daemon broadcasts The switch allows testing daemon-owned kernel execution while keeping local execution as the stable default. * Fix daemon execution order: launch kernel before queuing cells The daemon returns NoKernel when QueueCell is called without a running kernel, dropping the request. Fix by ensuring kernel is launched first, then queueing cells after it's ready. * Bump runtimed to 0.1.0-dev.3 and speed up dev builds - Bump version from 0.1.0-dev.2 to 0.1.0-dev.3 - cargo xtask build now uses debug mode for runtimed/runt-cli (~50s faster) - cargo xtask build-dmg/build-app still use release mode for distribution - cargo xtask install-daemon unchanged (always release for perf) * Update Cargo.lock for runtimed version bump

1. Socket path mismatch (Issue #1): Session.connect() now respects RUNTIMED_SOCKET_PATH env var, and test fixtures set it when spawning daemon in CI mode. 2. Nested timeouts (Issue #2): Remove outer 10ms timeout wrapping recv_frame_any() since it already has internal 100ms timeout. This prevents repeatedly canceling mid-read. 3. sync_rx not drained (Issue #3): Use try_send() instead of send().await for changes channel. If receiver isn't keeping up, skip the update rather than blocking. Python bindings keep sync_rx alive but don't consume it. 4. Parse failure semantics (Issue #4): When output_type is "error" but parsing fails, create an error Output to preserve success=false semantics. 5. CONDUCTOR_WORKSPACE_PATH (Safia's comment): Use env var as preferred repo root fallback before walking up parent directories.

* feat(runtimed-py): add PyO3 bindings for daemon client Add Python bindings for runtimed daemon operations: - DaemonClient: pool status, ping, list rooms, flush, shutdown - Session: connect to notebook room, start kernel, execute code - ExecutionResult/Output: structured output types Uses PyO3 0.28 with pyo3-async-runtimes for tokio integration. Note: Current Session.execute() uses QueueCell shortcut which bypasses automerge doc. See design gap notes for proper doc-based execution flow. * chore: add *.so to gitignore, remove stray architecture.md - Add *.so to root gitignore for Python extension modules - Remove contributing/architecture.md (lives on architectural-principles branch) * feat(runtimed-py): implement document-first execution Switch runtimed-py bindings to use ExecuteCell instead of deprecated QueueCell. The daemon now reads cell source from the automerge document, ensuring all connected clients see the same code being executed. Changes: - Add Cell class to expose cell info to Python - Add document operations: create_cell, set_source, get_cell, get_cells, delete_cell - Add execute_cell() using ExecuteCell (reads from doc) - Add run() convenience method (create + execute) - Fix sync task exit bug by storing sync_rx in SessionState - Add timeout to recv_frame_any() to prevent blocking - Add biased select! to prioritize commands over polling * test(runtimed-py): add daemon integration tests Add comprehensive integration tests for the document-first execution pattern. Tests cover: - Basic connectivity to daemon - Document operations (create/update/get/delete cells) - Cell execution via ExecuteCell (reading from automerge doc) - Multi-client synchronization (two sessions sharing a notebook) - Kernel lifecycle (start/interrupt/shutdown) - Output types (stdout/stderr/display_data) - Error handling Supports two modes: - Dev mode: uses existing daemon via `cargo xtask dev-daemon` - CI mode: spawns isolated daemon with log capture 24 tests covering the core document-first architecture. * ci(runtimed-py): add daemon integration tests to CI Add a new job to the build workflow that runs the runtimed-py integration tests. The job: - Builds runtimed binary and runtimed-py Python bindings - Runs 24 integration tests in CI mode (spawns isolated daemon) - Uploads test logs as artifacts for debugging Tests verify document-first execution, multi-client sync, kernel lifecycle, output capture, and error handling. * docs: add Python bindings documentation Comprehensive guide for the runtimed Python package covering: - Session API for code execution - DaemonClient for low-level operations - Document-first execution pattern - Multi-client scenarios - Result types (ExecutionResult, Output, Cell) - Sidecar launcher for rich output * docs(runtimed-py): add package README for PyPI * fix(runtimed-py): address code review findings 1. Socket path mismatch (Issue #1): Session.connect() now respects RUNTIMED_SOCKET_PATH env var, and test fixtures set it when spawning daemon in CI mode. 2. Nested timeouts (Issue #2): Remove outer 10ms timeout wrapping recv_frame_any() since it already has internal 100ms timeout. This prevents repeatedly canceling mid-read. 3. sync_rx not drained (Issue #3): Use try_send() instead of send().await for changes channel. If receiver isn't keeping up, skip the update rather than blocking. Python bindings keep sync_rx alive but don't consume it. 4. Parse failure semantics (Issue #4): When output_type is "error" but parsing fails, create an error Output to preserve success=false semantics. 5. CONDUCTOR_WORKSPACE_PATH (Safia's comment): Use env var as preferred repo root fallback before walking up parent directories.

rgbkrk added 3 commits February 11, 2026 13:55

feat: beginnings of a desktop notebook app

d755a04

Round III, let's go.

document usage of notebook

d2c2385

build all the UIs!

bd01108

rgbkrk merged commit f6e8bce into main Feb 11, 2026
3 checks passed

rgbkrk deleted the notebook-draft-1 branch February 11, 2026 22:23

rgbkrk mentioned this pull request Feb 20, 2026

Audit and overhaul Python environment management #167

Merged

5 tasks

rgbkrk added the daemon runtimed daemon, kernel management, sync server label Mar 14, 2026

rgbkrk mentioned this pull request Apr 15, 2026

feat(runtimed): llm_preview fields for blob-spilled stream and error outputs #1829

Merged

14 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: beginnings of a desktop notebook app#4

feat: beginnings of a desktop notebook app#4
rgbkrk merged 3 commits intomainfrom
notebook-draft-1

rgbkrk commented Feb 11, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

rgbkrk commented Feb 11, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant