Semantic Todoist Sync

Semantic Todoist Sync

Semantic Todoist Sync connects Obsidian notes with Todoist work. It supports grounded chat, note or selection to task capture, local references and deduplication, note/task synchronization, and a preview-first daily schedule. It is local-first: you choose the AI providers, models, and external services used for each workflow.

Install

• Install from Obsidian Community Plugins
• Download the latest release for a manual install

The plugin supports Obsidian 1.5.0 and later on desktop, iPhone, and iPad.

What it does

• Search and chat: Ask questions over relevant vault context and follow links back to source material. Internet Search and Deep Research are request-local, off by default, and retain supporting citations.
• Notes to Todoist: Turn a note, selected text, email, or prompt template into structured tasks, subtasks, and execution descriptions.
• Sync and deduplication: Keep note tasks and Todoist state aligned while preserving local references and distinct work.
• Schedule Today's Tasks: Preview a practical day, adjust it, then choose whether to apply or undo the Todoist changes.

Providers and model choice

The plugin supports OpenAI, Google Gemini, OpenRouter, and self-hosted Open WebUI connected to Ollama. Each workflow has a searchable provider/model picker and can use an optional fallback. Embedding indexes remain partitioned by their selected provider and model, so incompatible data is not mixed.

Embedding models remain user-selected and provider-scoped. The shipped default remains OpenAI text-embedding-3-large; choosing an OpenRouter embedding model such as openai/text-embedding-3-small creates a separate index identity rather than silently replacing an index built by another provider or model.

Provider defaults are conservative and adaptable rather than implicit model assumptions. OpenRouter starts with prompt-grounded JSON unless fresh capability evidence for the requested model or served route supports a native response schema. Open WebUI/Ollama starts with a prompt-grounded completed-object path unless authoritative discovery supports a native structured response. Exact Open WebUI model profiles may refine carrier, thinking, streaming, pacing, or output handling only after validation; stale, missing, or ambiguous metadata never enables an optional provider feature.

OpenRouter defaults to concurrency 10 and enforces a maximum of 16. Open WebUI defaults to one model-affine worker. Benchmark concurrency is a run-specific scheduling choice, not a model-quality claim.

All provider-visible generation, embedding, and native web-search inputs use a local preflight estimate below 16,000 tokens. The estimate is an admission and batching safeguard, not a provider-token guarantee. Provider-reported usage, when available, remains the post-dispatch observation.

Current Gemma4 exact-model profiles do not impose a hard model-specific output ceiling. Actual requests remain bounded by discovered context capacity, shared workflow safeguards, and any endpoint- or run-specific limits reported by the provider.

Quick setup

1. Open Settings > Semantic Todoist Sync and choose an AI provider.
2. Add the credential or configure an Open WebUI endpoint and sign-in details.
3. Add a Todoist token and run the connection check.
4. Choose models in the searchable operation settings and rebuild the semantic index from the setup screen.

Everyday use

Open the sidebar to use Ask, Tasks, Index, or Run. The command palette offers the same actions, plus note sync, email processing, scheduling, and undo.

For note task capture, use #STsync on main tasks and #STSubSync on subtasks. Required-action hashtags are configurable in settings; #todo is a shipped default example rather than a hard-coded retrieval rule. Configured action markers are mandatory coverage anchors, but they are not an exclusive source filter: clearly actionable unmarked content in the selected whole note is also examined. Normal note capture has no note-wide task cap; an explicitly selected single-task mode is the intentional exception.

Schedule Today's Tasks always opens a preview first. Todoist changes occur only after you select Apply, and the last applied schedule can be undone.

Output quality and privacy

• AI-assisted tasks and descriptions use the current request and relevant context. Only content needed for the configured workflow is sent to the selected provider.
• Local semantic indexes, task references, and scheduling memory stay on the device. Todoist receives task fields only when you use Todoist workflows.
• The plugin does not create a hosted account or require a single AI vendor.

Top five quality models from completed testing

This ranking retains only configurations marked completed with all expected rendered outputs in the published sanitized scorecard. It reflects the retained completed evidence for that fixed comparison, not a universal model rating, release-readiness statement, or claim about the current plugin state. Scores are rounded to one decimal from the published scorecard.

For the scoring approach and the complete sanitized aggregate, see the model-quality benchmark protocol and its published scorecard.

Optimizations in 0.8.21

• Fixed remaining selector usability regressions in primary/fallback model controls:
  • typed search input no longer gets overwritten by the selected label after user typing,
  • query text stays local to the search box until an explicit model is selected,
  • and minimum-3-character matching checks full provider/model text as before for dense OpenRouter/OpenAI/OpenWebUI catalogs.
• Confirmed selector lists are still grouped by provider and remain scrollable through long catalog lengths.

Optimizations in 0.8.20

• Settings model pickers now keep their behavior consistent across all primary/fallback and operation contexts:
  • both pickers are searchable from the selected provider list,
  • popup lists remain scrollable with dense long catalogs,
  • and query typing preserves the active filter when opening via typing or arrow-key navigation.
• Model selection remains provider-scoped in picker rows, so the configured primary/fallback selection set matches the active provider family used for each role.

Optimizations in 0.8.19

Task quality and scope safety

• Whole-note capture now treats configured required-action markers as mandatory coverage anchors rather than an exclusive filter. Clearly actionable unmarked sentences remain eligible, and the legacy per-note main-task cap is no longer applied. Context preflight and the configured per-task subtask limit still provide operational bounds.
• Each detected action is generated from an isolated source scope. Scope discriminators catch sibling-action substitutions and identical normalized outputs across different actions; one bounded recovery is allowed when the shared request budget remains, otherwise the conflict stays explicit.
• Task generation preserves exact actors, artifacts, conditions, timing, and explicit urgency. Deterministic priority correction and narrowly admitted structural repair can fix safe local issues without silently broadening the requested work.
• Task descriptions use compact singleton, task-local evidence ledgers with a cacheable shared prefix. They exclude stale or sibling work, avoid filler and source-title narration, and bind only facts actually stated. Unused canonical references can be removed or fixed deterministically without rewriting the model's prose; supplied context is guidance, not a fact-reproduction checklist.

Retrieval and request efficiency

• Local semantic routing adds an indexed lexical seed path, bounded top-K handle selection, revision-scoped caches, and source-thread corroboration so the plugin does not need to rescan every compatible chunk for each query.
• Eligible Open WebUI queries can use the configured local embedding model as a live ranking handle. Query vectors are identity-checked, deduplicated, cached, and batched within the same provider-input safeguards; indexed handles remain the fallback when a live query embedding is unavailable or ineligible.
• Generation, embeddings, and native web-search requests receive a local preflight below 16,000 estimated input tokens. Embedding batches split or stop before that boundary, while generation output allowances remain derived from the discovered context and operation rather than a Gemma-specific ceiling.
• Every logical generation lineage shares one budget of two provider dispatches: the initial request plus at most one additional request across carrier, transient, fallback, and workflow recovery paths. Deterministic parsing and normalization do not consume that budget.

Provider compatibility and reliability

• OpenRouter capabilities are accepted only from fresh model metadata; stale metadata falls back safely, and openrouter/free is treated as a dynamic served route rather than one fixed model identity. Free routes keep bounded rolling-rate admission, while paid-route limits remain provider-managed.
• Open WebUI uses conservative completed-object defaults and exact-model compatibility profiles where validation supports different thinking, streaming, schema, cooldown, or deadline behavior. Gemma4 E4B disables native thinking, uses completed delivery, and retains the provider-native schema for chat while task and description calls use prompt-grounded JSON.
• Open WebUI response handling now covers completed JSON, native NDJSON, and SSE. Error envelopes returned with HTTP 200 are classified and sanitized as provider failures instead of being accepted as model output or mislabeled as missing content, and oversized response bodies are rejected before parsing.
• Provider/model/operation-specific focus guidance supplements the shared schema contract for OpenAI, Gemini, OpenRouter, and Open WebUI without making unsupported transport assumptions for every provider.
• Global task-insertion and immediate-sync settings are authoritative over prompt-template defaults, preventing a template from enabling note or Todoist writes that the user disabled globally.

Links

• Changelog
• License — GNU General Public License v3.0