Command Center for Obsidian

Command Center is a personal operational OS for Obsidian — a local-first AI multi-agent orchestrator that transforms your vault into an intelligent workspace. It combines interview-driven onboarding, a 13-provider model routing layer, a local Pi ReAct engine with Orchestrator–Worker loops, hybrid vault RAG (BM25 + semantic retrieval), persistent agent memory, native Markdown/Canvas workflows with Bases queue integration, voice input/dictation/live transcription, a daily operations engine, and headless automation — all without imposing a productivity framework on your notes.

[!IMPORTANT] Command Center can modify vault files when you approve or run mutating operations. Keep backups, review destructive action cards, and test workflows on non-critical notes first.

Why Command Center?

Most AI integrations add a chat box. Command Center adds an operational layer:

• Framework-agnostic by design — no required PARA, GTD, Johnny.Decimal, folder taxonomy, note schema, or life-management methodology. The setup interview learns the structures already used in your vault.
• Metacognitive Partner, not prescriptive organizer — two-stage Socratic Triage first learns your goals, constraints, language, and personal definition of efficiency, then negotiates topology observations as neutral evidence. It preserves intentional friction and proposes capabilities only with consent.
• Interview-driven configuration — embedded dashboard discovery collects topology, life map, capacity rules, triage policy, focus constraints, and writing style; confirmation and synthesis then generate only the assets you approve.
• Token-efficient stationary indexing — protected _index.md files describe folder purpose and direct-child contents so agents can route work without repeatedly scanning the whole vault.
• Built-in Obsidian Secrets — API keys are managed through Obsidian's native Secret Storage, so they persist with your vault instead of a custom encrypted file. Keys are edited from a dedicated settings modal, and keyed cloud providers remain available whenever their secrets are present.
• Local-first options — Pi, Ollama, LM Studio, custom OpenAI-compatible endpoints, local embeddings, and deterministic fallback retrieval support private or disconnected workflows.
• Human control at mutation boundaries — destructive and bulk changes pause on approval cards with collapsible, syntax-colored diff previews.

Contents

• Feature overview
• Architecture
• Installation
• First-run onboarding
• Configuration
• Everyday usage
• Workflows and Bases
• Voice and automation
• Privacy and security model
• Development
• Future improvements
• Quality, security, and release controls
• Community and support
• License and attribution

Feature overview

Multi-Provider v2.0

Command Center exposes one dispatch layer across 13 providers:

Provider	Type	Notes
Pi Daemon	Local companion	Serialized Pi JSONL RPC; keyless at the plugin boundary
OpenAI	Cloud	GPT/o-series, vision, embeddings
Anthropic	Cloud	Claude, tools, prompt caching
Google Gemini	Cloud	Multimodal, long context, cached-content support
OpenRouter	Cloud gateway	Multi-model OpenAI-compatible routing
Ollama	Local	Local chat, keep-alive lifecycle controls, and optional bearer authentication
Groq	Cloud	Low-latency inference and transcription-compatible routing
DeepInfra	Cloud	Hosted open-weight models
Mistral AI	Cloud	Mistral, Codestral, and Voxtral families with native STT/TTS (/v1/audio/transcriptions, /v1/audio/speech).
Cohere	Cloud	Command models for RAG with native STT (/v2/audio/transcriptions).
LM Studio	Local	Dynamic native model resolution, resource-aware JIT loading, OpenAI-compatible inference, and optional bearer authentication
xAI (Grok)	Cloud	Grok models with vision, tools, native STT (/v1/stt), and TTS (/v1/tts).
Custom Endpoint	Local or remote	User-defined OpenAI-compatible service with optional bearer authentication

Routing classifies work as coding, vision, reading, reasoning, or fast. Capability checks prevent invalid model selection; optional exponential moving averages optimize initial routes for latency, cost, or a balanced objective. Recovery remains reliability-first: authentication and invalid-request failures fail or fall through immediately, while rate limits, network errors, timeouts, and server errors use isolated circuit breakers, bounded backoff, and a configurable multi-tier fallback chain.

Additional provider capabilities include:

• Background model discovery with persistent catalogs and static-registry fallback — live models auto-fetch from each enabled provider's /models endpoint on startup, covering chat, STT, and TTS model resolution
• Anthropic prompt caching and Gemini cached-content bookkeeping
• Multimodal image preprocessing for vault attachments and Canvas file nodes
• Structured-output repair only at model-authored JSON boundaries
• Local model pre-warm, TTL/keep-alive, best-effort eviction, and auto-download of missing models
• LM Studio model download via POST /api/v1/models/download with progress tracking
• Dynamic LM Studio model resolution through /api/v1/models: reuse a loaded primary LLM or select the smallest downloaded non-draft conversational model, excluding embedding and speculative draft models
• Optional secure-vault bearer tokens for LM Studio Require Authentication, authenticated Ollama proxies, and custom OpenAI-compatible endpoints; unauthenticated local operation remains supported
• Cloud-bound payload scrubbing of local-only lifecycle fields

Multi-Agent ReAct engine

The ReAct runtime follows an Orchestrator → Worker → Observation → Correction loop:

1. The orchestrator reasons about the objective and selects one or more workers.
2. Independent workers are planned in parallel and transported safely over serialized Pi RPC.
3. Observations return to the orchestrator with evaluation scorecards.
4. Validation detects incomplete, circular, hedged, conflicting, or error-heavy output.
5. The loop corrects, re-routes, or synthesizes a final answer.

Built-in worker profiles cover orchestration, retrieval, summarization, and structural editing. ReAct-capable profiles (react-orchestrator, react-analyst) extend these with iterative reason-act-observe loops. Five standard agent roles—Orchestrator, Triage, Indexer, Health, and System Architect—bind each operational responsibility to a compute tier and least-privilege tool ceiling, and the runtime can create constrained custom roles without granting tools outside the parent worker's ceiling.

Role vocabulary (three layers)

The codebase uses three distinct, layered vocabularies — they are intentionally not collapsed, because each keys a different table:

• WorkerProfileName (src/types.ts) — the 4 static prompt+token configs in src/workers/: orchestrator, retriever, summarizer, editor. Smallest and most stable.
• AgentWorkerProfile (src/execution/ExecutionRouter.ts) — a superset adding react-orchestrator and react-analyst, the ReAct-capable profiles that have no static prompt entry but declare an execution modality (text/embeddings/…).
• StandardAgentRole (src/engine/AgentTypes.ts) — the 5 operational roles above; each maps to a compute tier + a worker profile + a TaskType via AGENT_TAXONOMY. Task.workerRole reuses this union.

The pi-daemon string is a sentinel, not a profile: command-palette local tasks set workerProfile: 'pi-daemon' to route directly to the local Pi daemon via router.routeDirect.

Operational safeguards include:

• Same-cycle target conflict detection and normalized FIFO file locks
• Circuit breakers, retry/backoff, timeout recovery, deadlock detection, and safe-state rollback
• Four-dimension evaluation: completeness, relevance, specificity, and correctness
• Cooperative Debug / Step Mode that pauses only at safe cycle boundaries
• A fixed 50-row ReAct monitor with All, Actions & corrections, and Errors filters
• Full-fidelity session replay and Markdown audit export

Hybrid Vault RAG and persistent memory

Command Center can ground model calls in vault content without rebuilding the index on every query:

• Markdown is chunked into semantic 300–500-word units aligned to headings, paragraphs, and sentences.
• Chunk metadata retains path, H1/H2 hierarchy, normalized wikilinks, and inclusive line ranges.
• Path + mtime + size keys skip unchanged reads and embeddings.
• BM25 lexical ranking and cosine semantic ranking are merged through weighted reciprocal rank fusion (RRF).
• Folder scopes constrain retrieval, and cited snippets remain inside a hard prompt budget.
• Unavailable embedding services degrade to deterministic local term-frequency vectors.

Persistent agent memory stores facts, preferences, entities, and session summaries in vault-native state. Semantic duplicate updates, thematic session hubs, threshold-aware pruning, and bounded prompt injection keep memory useful without allowing it to grow without control.

Metacognitive Partner and semantic memory

Command Center acts as a Metacognitive Partner: it helps users examine and negotiate how their own system works rather than grading it against a generic productivity framework. Discovery follows two deliberate stages:

1. Contextual baseline — one focused Socratic question at a time establishes goals, constraints, working context, preferred cognitive style, and the user's subjective definition of success and efficiency.
2. Topographical negotiation — only after that baseline exists does Command Center introduce read-only vault observations. Patterns are treated as hypotheses; intentional exceptions and useful friction are preserved, and automation, semantic linking, or multi-agent synthesis are offered as optional capability expansions with explicit tradeoffs.

The metacognition layer builds supporting local context without reorganizing or rewriting user notes:

• TopographySweep uses Obsidian's TAbstractFile, TFile, TFolder, MetadataCache, and getAllTags APIs to map folders, tag frequencies, links, and hub/MOC candidates.
• The sweep runs silently and cooperatively, excludes .obsidian and .trash, and writes only .obsidian/plugins/command-center/vault_topography.json.
• Logic Discovery onboarding is consent-led and treats topology as neutral evidence requiring user confirmation. Confirmed preferences and the transcript are stored in user_logic_profile.json.
• Header-aware Markdown chunking preserves ##/### boundaries, source lines, frontmatter tags/aliases, and outbound [[wikilinks]].
• The semantic database schema separates documents, chunks, and vectors. With an injected desktop SQLite-VSS driver, document replacement is transactional and nearest-neighbor search is persisted locally; the distributed build otherwise uses a process-lifetime in-memory index.
• Dialectic RAG requests only the embeddings modality through the Native Auto-Router and Python execution boundary. The shipped Python worker is a secure transport stub and reports that no embedding backend is configured; an integrated backend must return normalized, dimension-validated vectors before they can be stored.

Native workflows and Bases queues

Workflows are vault-native rather than hidden in a remote service:

• Parse Markdown frontmatter or JSON Canvas graphs into validated DAGs
• Collect typed text, dropdown, and toggle inputs in native Obsidian modals
• Execute independent steps in parallel topological tiers
• Evaluate constrained conditions over inputs and earlier step results—without eval
• Interpolate {{inputs.*}} and {{steps.*.result}}
• Route each step through a provider or Pi policy
• Export Markdown workflows to deterministic, executable Canvas layouts

The Command Center Queue integrates with Obsidian Bases. It consumes native evaluated results, preserves filters/formulas/sorts/limits, excludes terminal notes, supports selection and bounded concurrency, and writes agent_status, score, and timestamp fields through processFrontMatter() so active Bases views refresh in real time.

Daily operations and stationary indexes

The interview becomes the source of truth for daily operations:

• Morning capacity evaluation, inbox proposals, and daily-note assembly
• Midday text or voice updates under a stable log heading
• Evening review with explicit complete, rollover, discard, or leave decisions
• Optional Silent Daily Startup that assembles and opens the note while leaving inbox mutations unapproved
• Collision-safe inbox move/archive/extract/delete proposals
• Frog/aging audits and configurable capacity rules
• Protected _index.md manifests maintained from direct-child scans

Stationary indexes contain purpose, scope, summary, and status metadata. Compact purpose headers allow routing to the correct folder before deeper retrieval, reducing full-vault reads and prompt waste.

Dashboard Logic Discovery

The full-page Command Center Dashboard is the single operational interface for Socratic vault discovery, onboarding, agent monitoring, queue control, approvals, and daily operations. Discovery is a dashboard mode—not a separate deck, pane, or modal. It begins with a contextual baseline before introducing read-only TopographySweep evidence. Topology remains supporting evidence and is never promoted into a rule without user confirmation.

Logic Discovery uses bounded generation and disables model reasoning where supported so the dashboard presents one concise visible question at a time. With LM Studio enabled, Command Center discovers native catalog state through /api/v1/models, prefers an already loaded primary conversational model, or JIT-loads the smallest suitable downloaded model before calling /v1/chat/completions.

Open Command Center from the ribbon or run Command Center: Start Setup / Onboarding Interview. Both routes use the same full-page dashboard.

MCP (Model Context Protocol) tool discovery

Command Center now supports the Model Context Protocol for discovering and executing external tools from MCP servers. Add MCP server URLs in settings to make their tools available to the LLM during inference:

• JSON-RPC 2.0 transport over HTTP/SSE
• Dynamic tool discovery via tools/list — tools are wrapped as ToolDefinitions
• Error isolation — one MCP server failing doesn't affect others
• LM Studio MCP — LM Studio can be configured as an MCP host at /api/v1/mcp

Normalized execution and Shadow-Clone diagnostics

Standard and Python-backed agent work crosses a mandatory execution boundary:

• NativeAutoRouter reads the shipped model_matrix.json and applies the global 1–10 quality/cost depth to text, image, audio, video, and embedding intents.
• Orchestrator, summarizer, editor, retrieval, and ReAct worker profiles declare an explicit modality before provider/model resolution.
• Python workers use isolated JSON-RPC 2.0 subprocesses over stdin/stdout with shell: false, bounded output, cancellation, timeouts, cleanup, and circuit breaking. Credentials are never placed in argv or environment variables.
• DataNormalizer is the trust boundary for provider responses, Python results, intermediate observations, and multi-agent merges. It sanitizes tracebacks, stderr, malformed JSON, control characters, and oversized output before UI or vault use.
• Locked or failed keyed routes retain local-first behavior through provider usability checks and circuit-breaker fallback.

Run Command Center: Run Shadow-Clone Diagnostics from the command palette to verify credential-memory wiping, current slider/matrix routing, fail-safe local routing, and Python-output sanitization. The harness uses in-memory fixtures and prints a sanitized report to the developer console; it does not modify user notes.

OpenRouter web search

When enabled in settings, Command Center includes a server-side web search tool in requests routed through OpenRouter. The model can invoke web_search_call to pull live information from the web, with results and citations returned directly in the response. Controlled by the webSearchEnabled setting toggle.

Native chat and context

The right-sidebar chat supports Quick, ReAct, and Workflow modes with:

• Provider/model status and streamed Markdown responses
• Bounded multi-turn history
• Active editor selection, @Note, @path, and .base context
• Dismissible active/recent-note suggestion pills
• Inline collapsible ReAct traces
• Dashboard handoff for approval/rejection of destructive tools
• Tail-aware auto-scroll and lifecycle-safe cancellation
• Message actions: copy to clipboard, delete, and read-aloud buttons with SVG icons
• Code block copy: hover-revealed copy button on every rendered code block
• Blinking streaming cursor (▊) on pending assistant messages
• Hover-revealed action bar: timestamp, copy, delete, and read-aloud fade in on bubble hover
• Scroll-to-bottom button: appears when chat history is scrolled up
• New conversation and Markdown export with tagged frontmatter

Capability Registry — unified agent tool-calling

Command Center now includes a Capability Registry — a central, discoverable surface for every instrument the agent can invoke. Instead of the orchestrator pre-selecting which tools to use, the model can autonomously reason about which capability serves each task:

• Unified registry: Vault tools (read, write, search, list), web search, MCP-discovered tools, and agent worker profiles all register in one place.
• User-configurable: Each capability can be enabled or disabled from Settings, grouped by category (Search, File Operations, Media, Time, Memory, System, MCP, Agent).
• Autonomous mode: The model can decide to call a tool on its own, or only when the user explicitly requests it via @-command aliases.
• Alias-aware: Capabilities expose @-command aliases (e.g., @vault, @websearch, @composer, @memory) so users can invoke tools explicitly.
• Execution modes: always (always included in context), autonomous (model may decide), explicit (only on user request).
• Confirmation policies: never, on-threshold, or always for destructive operations.
• Event-driven UI: The registry emits events when capabilities are registered, enabled, or disabled — the Settings UI and dashboard widgets react in real time.
• System prompt injection: describeEnabled() produces a compact inventory of available capabilities for the model's context window.

Project Mode — isolated AI workspaces

Projects are focused AI workspaces with isolated chat history, per-project model configuration, and scoped context sources:

• Vault-native storage: Each project is a .md file under .command-center/projects/ with YAML frontmatter — no hidden databases or external services.
• Per-project model: Override the global provider/model for a specific project.
• Custom system prompt: Assign a system prompt to shape the agent's behavior for that project.
• File scoping: Inclusion and exclusion patterns control which notes are visible to the project's agent.
• Web and YouTube context: Pre-load web page URLs and YouTube video transcripts as always-available context.
• Isolated chat history: Conversations in one project never bleed into another.
• Archiving: Archive projects to hide them from the active list while preserving data.
• Usage tracking: Automatic last-used timestamps and conversation counts for sorting.

Inline Composer — surgical text editing with diff preview

The composer provides a three-stage fuzzy matching engine for precise text replacement:

1. Exact match — after line-ending normalization (CRLF → LF)
2. Fuzzy match — NFKC normalization, smart quotes, special dashes, non-breaking spaces
3. Trimmed match — retry after stripping trailing newlines

Additional capabilities:

• BOM handling: UTF-8 BOM is preserved through edit cycles.
• LCS-based diff: Longest-common-subsequence diff computation with line-level granularity.
• Multi-operation editing: applyOperations() applies a sequence of insert, update, replace, and delete operations in order.
• Diff statistics: Additions, deletions, and unchanged line counts for progress reporting.

@-Mention Typeahead — inline vault references

The typeahead engine provides real-time suggestions as you type @ in the editor:

• Multi-source: Searches notes, folders, tags, and capabilities simultaneously.
• Categorized results: Results grouped by source type with clear labels and descriptions.
• Keyboard navigation: Arrow keys, Enter/Tab to select, Escape to dismiss.
• Vault caching: 30-second TTL cache prevents repeated vault scans.
• Capability aliases: @vault, @websearch, @composer, @memory resolve to their corresponding capabilities.
• Wikilink insertion: Selected notes are inserted as [[wikilink]] references.

User-Managed System Prompts

System prompts are stored as vault-native Markdown files with YAML frontmatter:

• CRUD management: Create, edit, delete, and browse prompts from the vault.
• Variable substitution: {{vault}}, {{date}}, {{time}}, {{user}}, {{style}}, {{memory}} are resolved at render time.
• Custom resolvers: Override variable resolution per-project or per-chat-mode.
• Default prompt: A sensible default prompt is created on first launch.
• Category filtering: Organize prompts by category for quick access.

User Memory Manager — explicit "remember this"

Builds on the persistent agent memory store to provide user-facing memory operations:

• Explicit memory: remember() processes natural-language "remember that" commands.
• Auto-extraction: extractFromTurn() detects "I prefer", "I am", "remember that" patterns in conversation turns.
• Profile building: buildProfile() aggregates stored facts into a structured UserMemoryProfile with name, style, expertise, and goals.
• Contextual recall: recall() searches memories by relevance to a query and returns formatted Markdown.
• System prompt injection: injectMemoryPrompt() produces a bounded memory context block for the model.

Architecture

Obsidian desktop
├── Command Center dashboard
│   ├── task queue, status, history, and frame-batched streaming
│   └── fixed-pool ReAct monitor, replay, export, and debug stepping
├── Right-sidebar chat
│   ├── Quick → ConversationManager → ProviderDispatcher
│   ├── ReAct → PiAgentDaemon → Orchestrator/Workers/Tools
│   └── Workflow → WorkflowEngine → provider or Pi routes
├── Interview-derived operations
│   ├── ConfigManager + generated style guide
│   ├── FolderIndexer → protected _index.md manifests
│   └── DailyEngine + InboxTriager + CapacityEngine
├── Metacognition and knowledge layer
│   ├── TopographySweep + dashboard LogicDiscovery → localized topology/profile JSON
│   ├── ChunkingEngine → H2/H3 chunks + tags/aliases/wikilinks
│   ├── DialecticRAG → normalized embedding ingestion → memory / injected SQLite-VSS
│   ├── HybridRetriever → BM25 + embeddings + weighted RRF
│   └── AgentMemoryStore/ReActMemoryBank → bounded persistent memory
├── Workflow layer
│   ├── Markdown/Canvas parser → validated DAG tiers
│   ├── Bases queue → bounded target batches
│   └── frontmatter state sync → live Bases refresh
├── Execution layer
│   ├── NativeAutoRouter → model_matrix.json + global depth 1–10
│   ├── ExecutionRouter → explicit worker modalities + secure credentials
│   ├── PythonWorkerTransport → bounded JSON-RPC subprocesses
│   └── DataNormalizer → sanitized results, observations, and merges
└── Provider layer
    ├── ModelRouter/ProviderDispatcher → capability, fallback, isolated circuits
    ├── cloud adapters → OpenAI, Anthropic, Gemini, etc.
    └── local adapters → Pi, Ollama, LM Studio, custom endpoint

Important boundaries

• Obsidian-native I/O: vault files are read and changed through Obsidian APIs wherever possible.
• Shared write coordination: UI, CLI, daily services, indexers, workflows, and tools use one normalized per-vault lock namespace.
• Serialized Pi transport: Pi events do not identify an originating prompt, so one active RPC prompt is correlated through agent_end and agent_settled.
• Raw-byte framing: JSONL is split on byte 0x0A before UTF-8 decoding, preserving split code points, CRLF, U+2028/U+2029, bursts, and final unterminated frames.
• Bounded UI work: model deltas and trace events are coalesced with requestAnimationFrame; trace history and visible rows are capped.

Installation

Requirements

• Obsidian 1.13.0 or newer
• Desktop Obsidian (the plugin is desktop-only)
• A provider configured in Command Center settings
• Optional: Node.js 20+ and Pi 0.83.0+ for local Pi/ReAct execution

Manual installation from a GitHub release

1. Download the release assets or the packaged command-center directory.

2. Create this folder inside your vault:

   <your-vault>/.obsidian/plugins/command-center/
   

3. Copy the plugin assets from release/command-center/:

   command-center/
   ├── main.js
   ├── manifest.json
   └── styles.css
   

   Repository-level LICENSE and ATTRIBUTIONS.md remain published alongside the source and GitHub release documentation.

4. Restart Obsidian or reload community plugins.

5. Open Settings → Community plugins and enable Command Center.

6. Run Command Center: Start Setup / Onboarding Interview from the command palette.

Optional Pi installation

Pi powers local palette tasks and the full multi-agent ReAct path. Install it separately:

npm install -g @earendil-works/pi-coding-agent
pi --version

Command Center auto-detects common global npm locations. On Windows it resolves pi.cmd to Pi's JavaScript CLI and launches it with the real Node executable to avoid Electron/Node wrapper issues. You can override the detected path in Settings → Command Center → Core Configuration.

Pi is an external MIT-licensed companion and is not bundled into this plugin.

First-run onboarding

Launch onboarding in the full-page dashboard from either the Command Center ribbon action or Command Center: Start Setup / Onboarding Interview. Discovery, confirmation, and synthesis stay in the central workspace instead of opening a separate setup modal.

The six configuration phases are:

1. Topology — inboxes, daily notes, managed folders, and vault structure
2. Life map — domains, projects, time horizons, and completion definitions
3. Capacity — tracked metrics and rules that scale daily commitments
4. Triage — move, archive, extraction, deletion, and aging policies
5. Focus — priority caps, frog rules, quick wins, and task conventions
6. Style — writing voice, agent persona, vocabulary, layouts, and reflections

A confirmation/synthesis stage then previews 2–4 templates and 2–3 workflows. Nothing is generated until you explicitly select and approve it.

The interview writes validated assets under .command-center/, including:

.command-center/
├── config.json
├── style-guide.md
├── templates/
└── workflows/

These are local runtime files and must not be published with the plugin. The repository includes a non-secret reference at docs/config.example.json.

Onboarding security rule

Do not enter API keys, passwords, tokens, URLs, hosts, ports, or endpoint details into the interview. Secret-like input is rejected locally. Configure provider credentials and endpoints only in Settings → Command Center.

To start over, run Command Center: Reset / Re-Initialize Vault Configuration. Generated configuration and style files are moved through Obsidian's trash flow before onboarding restarts.

Configuration

Open Settings → Command Center. The settings UI is organized into six sections.

1. Core Configuration

Configure the active profile, token limits, Pi path, daemon startup, memory limits, Base batch concurrency, and Silent Daily Startup. Pi detection and status diagnostics are available here.

2. Accessibility & Speech

Configure text-to-speech enablement, speaking voice, speaking rate, speech-to-text enablement, transcription provider preferences, and automatic read-aloud behavior here. Chat and voice recording use the same speech settings.

Speech-to-text models are per-provider (STT model IDs are not portable across providers — openai/gpt-4o-mini-transcribe is an OpenRouter routing slug, grok-stt is xAI, whisper-1 is OpenAI). Set the slug each provider accepts in the per-provider model fields; a blank entry uses the provider's built-in default.

Text-to-speech can use the browser's built-in speech engine (default) or route through a provider's /audio/speech (or xAI /v1/tts) endpoint for higher-quality voices. Pick the engine in Text-to-speech engine; set a per-provider TTS model and voice id when using a provider engine.

3. Provider Credentials

Each provider has a collapsible card for enablement, endpoint configuration, health checks, and model refresh. API keys are not exposed through ordinary settings fields.

Select Manage API Keys to open the built-in secrets editor. Provider secrets are stored in Obsidian Secret Storage under the Command Center namespace, so they persist with your vault instead of a custom encrypted file. Existing secrets can be replaced or removed, but they are not revealed back into the UI.

Authentication metadata distinguishes required, optional, and unsupported credentials. LM Studio's Require Authentication token, authenticated Ollama proxies, and custom OpenAI-compatible bearer tokens all use Obsidian Secret Storage for persistence without making a key mandatory for ordinary local operation. Tokens are applied consistently to inference, streaming, model discovery, health checks, transcription, and local model lifecycle requests.

Credentials are resolved only at request time. Provider secrets stay in Obsidian's secret store, while interviews, generated workflows, CLI/URI arguments, subprocess argv/environment, logs, and repository examples continue to exclude credentials.

4. Task Routing Matrix

Assign a provider/model pair to each task class:

Task class	Typical use
coding	Code generation, refactoring, technical edits
vision	Image and Canvas attachment analysis
reading	Long documents, synthesis, extraction
reasoning	Planning, analysis, orchestration
fast	Classification and low-latency transforms

Live-discovered models appear with a network indicator. If discovery fails, the static registry remains available.

5. Fallback Pipeline

Enable or disable fallback, then add, remove, and reorder providers. Permanent request/schema errors fail fast; transient failures use backoff and reliability-ranked alternatives without allowing cost optimization to weaken recovery.

6. Health Dashboard

Review provider state, test one provider, refresh all providers, and inspect actionable errors such as a missing Pi binary or unreachable local endpoint.

Everyday usage

Dashboard

Open Command Center from the ribbon or command palette. The dashboard includes:

• Embedded Socratic discovery and onboarding
• Daemon Start / Stop / Restart controls
• Pending, running, completed, and failed queue counts
• Provider-normalized orchestrator output and per-task live output
• Task history, ReAct filters, debug stepping, and session export
• Dashboard-owned approval cards for destructive or bulk mutations
• Daily-cycle controls and consolidated silent-start summaries
• Per-vault widget ordering, sizing, collapse/visibility controls, and responsive layout persistence
• Review-before-send dictation plus opt-in audio cues, speech-to-text, and AI read-aloud

Chat panel

Run Command Center: Open Chat Panel and choose:

• Quick for provider-routed conversational work
• ReAct for orchestrated multi-agent execution
• Workflow to run a Markdown, Canvas, or Base-backed workflow

Use @Note Name, @folder/note.md, an active editor selection, or a .base reference to attach vault context. Suggested recent/active notes appear as dismissible pills; only retained pills are sent.

Local task

Open a Markdown note and run Execute agent task on current note. This command explicitly routes through the local Pi daemon and starts it automatically when possible.

Workflows and Bases

Markdown workflow concepts

A workflow defines typed inputs and steps. Every synthesized step declares an assigned agent, required compute tier, fallback policy, and action type. Dependencies form a DAG; unknown dependencies and cycles are rejected before execution.

Conceptual example:

---
workflow:
  name: Review incoming note
inputs:
  tone:
    type: dropdown
    options: [concise, detailed]
steps:
  - id: inspect
    assigned_agent: researcher
    required_tier: tier1_local
    fallback_policy: configured
    action_type: read
    prompt: "Inspect {{inputs.targetPath}} using a {{inputs.tone}} style."
  - id: summarize
    dependsOn: [inspect]
    assigned_agent: writer
    required_tier: tier2_reasoning
    fallback_policy: configured
    action_type: write
    prompt: "Summarize: {{steps.inspect.result}}"
---

The exact accepted shape is validated by the native parser and may include input defaults, conditions, provider/Pi routing, and output metadata.

Canvas workflows

Open a Markdown workflow and run Export Active Workflow to Canvas. Parallel steps share a tier column, dependencies become directed edges, and executable step metadata is retained in text nodes.

Bases queue execution

A standalone .base file can reference a workflow. Open it and run Execute Workflow on Current Base Queue. Choose concurrency from 1–10 and an optional partial-run limit. Command Center:

1. Resolves Obsidian's evaluated Base entries
2. Excludes completed/failed notes
3. Runs bounded independent target contexts
4. Awaits frontmatter updates between tiers
5. Refreshes and re-queries membership before continuing

This allows a Base to function as a live, self-draining agent queue.

Voice and automation

Voice input

The chat microphone and Command Center: Quick Voice Prompt use browser-native MediaRecorder:

• In-memory audio assembly
• Live timer and level meter (chat, voice prompt modal, and dashboard dictate button)
• Deterministic microphone-track cleanup
• OpenAI-compatible multipart transcription (per-provider model resolution — see below)
• Retry only for transient network/408/429/5xx failures
• Mic button disabled when no STT provider is configured, with tooltip feedback
• Spoken @ mentions and active-selection context resolution
• Contextual delivery — the Quick Voice Prompt routes transcribed text by focus: into the active note at the cursor when a note is in focus, or into the chat input field (for review before send) when the chat panel is in focus or no note editor is active

The transcription fallback chain tries each enabled STT-capable provider in order (LM Studio, Ollama, Groq, OpenAI, DeepInfra, Mistral, OpenRouter, xAI, Cohere, Custom). Each provider uses its own model slug — set per-provider in Settings → Accessibility & Speech, or leave blank for the built-in default (whisper-1 for OpenAI, grok-stt for xAI, whisper-large-v3 for Groq, openai/whisper-large-v3-turbo for DeepInfra, voxtral-mini-latest for Mistral, openai/whisper-large-v3 for OpenRouter, cohere-transcribe-03-2026 for Cohere). A global model is no longer broadcast to every provider, so a foreign slug can no longer break a provider it wasn't meant for.

Audio is sent only to the transcription endpoint configured in your local settings. Review that provider's privacy policy before use.

Text-to-speech

Spoken output defaults to the browser's built-in speechSynthesis engine. When Text-to-speech engine is set to a provider (or Auto), Command Center routes the text through that provider's TTS endpoint instead:

• xAI: POST /v1/tts (native, model grok-tts)
• OpenAI: POST /v1/audio/speech (model gpt-4o-mini-tts)
• OpenRouter: POST /api/v1/audio/speech (routed slugs like openai/tts-1)
• Mistral: POST /v1/audio/speech (Voxtral TTS)

The returned audio plays through a hidden <audio> element. If the provider request fails, Command Center falls back to the browser engine so spoken output is never silently dropped. Set a per-provider TTS model and voice id (e.g. alloy, nova, coral) in Settings → Accessibility & Speech.

Native Obsidian CLI

On Obsidian versions exposing native CLI registration:

command-center:morning
command-center:workflow
command-center:indexes

These handlers invoke service layers directly without opening views. They return structured JSON and use non-zero failure semantics. Morning automation assembles the configured note but never auto-approves inbox mutations.

Example shape:

obsidian command-center:morning --metrics '{"available_minutes":120}'
obsidian command-center:workflow --path '.command-center/workflows/example.md' --inputs '{}'
obsidian command-center:indexes

Consult your installed Obsidian CLI's command syntax because host invocation details may vary by version.

URI fallback

For hosts without native CLI registration:

obsidian://command-center?operation=morning
obsidian://command-center?operation=indexes

The URI and CLI boundaries reject credential arguments, unsafe vault paths, malformed/oversized JSON, and execution before onboarding is complete.

Privacy and security model

Command Center is local software, but it uses the network when you select a cloud model or transcription provider, refresh a remote model catalog, or connect to a network-hosted custom endpoint. Supported remote services include OpenAI, Anthropic, Google Gemini, OpenRouter, Groq, DeepInfra, Mistral AI, and Cohere; requests are used only to provide the model, embedding, discovery, or transcription feature you invoke.

The desktop-only plugin also accesses files outside the vault in two explicit cases: it launches the separately installed Pi CLI as a child process, and it may read image attachments referenced by absolute paths so they can be sent to your selected vision provider. Ordinary note, configuration, memory, workflow, and index operations use Obsidian's vault APIs.

No client-side telemetry or advertising is included. Command Center does not self-update; installation and updates are handled by Obsidian and GitHub releases.

Data that may be sent to a model

Depending on the mode and context you approve:

• Your prompt and bounded conversation history
• Selected note text or resolved @ attachments
• Retrieved and cited vault snippets
• Relevant persistent memory entries
• Images explicitly referenced by the prompt
• Workflow inputs and prior step results

Use local Pi/Ollama/LM Studio/custom endpoints when content must remain on your network, and verify those services independently.

Local safeguards

• Secrets are entered through Obsidian's built-in Secret Storage UI from the Command Center settings tab.
• Provider secrets persist in Obsidian Secret Storage under the command-center namespace.
• Existing secrets are managed in place: replace them, remove them, or leave them unchanged.
• Interview, config generation, CLI/URI arguments, subprocess argv/environment, and examples prohibit credentials.
• Tool paths reject traversal, absolute paths, NUL bytes, and unsafe characters.
• Existing-note overwrites and bulk/destructive operations can require explicit approval.
• Shared FIFO locks coordinate all mutation paths.
• Prompt context, memory, traces, histories, Python subprocess output, and UI rows are bounded.
• All Python JSON-RPC responses, stderr, tracebacks, malformed output, and multi-agent merges pass through DataNormalizer.
• .gitignore excludes runtime config, memory, audit traces, topology, generated assets, logs, environment files, and release/build outputs.
• scripts/sanitize-repo.mjs scans repository and release content for common secrets, private keys, absolute local paths, private IPs, NULs, and tracked runtime state.

Repository sanitization

Run before committing or publishing:

npm run sanitize          # tracked + untracked public files
npm run sanitize:staged   # staged Git blobs; suitable for pre-commit
npm run sanitize:release  # repository plus current build/release assets

npm run package automatically performs the release scan after packaging. CI and release workflows also run the sanitizer.

Development

Obsidian plugin development baseline

This repository follows the official Obsidian plugin documentation for lifecycle, workspace, settings, commands, modals, events, vault access, and submission requirements. When changing code, prefer Obsidian-native APIs over custom replacements, clean up event handlers on unload, and keep desktop-only code paths explicit where the plugin depends on them.

Prerequisites

• Node.js 20, 22, or 24
• npm
• A desktop Obsidian vault for manual integration testing

Setup

git clone https://github.com/scrunchds/Command-Center.git
cd command-center
npm ci
npm run dev

npm run dev starts esbuild watch mode. Reload Obsidian after bundle changes.

Commands

Command	Purpose
npm run typecheck	Strict TypeScript check (including security, metacognition, execution, and diagnostic layers)
npm run lint	Zero-warning ESLint gate
npm run test	167 core + 153 ReAct + 22 provider = 342 total
npm run benchmark	Produce the standardized 10-metric report
npm run benchmark:check	Enforce the 25% core regression threshold
npm run sanitize	Scan public repository files for PII/secrets/runtime data
npm run sanitize:staged	Scan staged blobs for pre-commit use
npm run build	Typecheck and build minified production JS/CSS
npm run package	Build, create a clean release folder, and sanitize it
npx obsidian-plugin-validator .	Run community-submission manifest and source checks

Release output

npm run package recreates release/command-center/ from scratch and permits exactly:

main.js
manifest.json
styles.css

License and attribution documents remain at repository level. Restricting the installable directory to Obsidian's three production files prevents stale development or documentation files from leaking into the plugin package.

Future improvements (not yet implemented)

Chat conversation management — delete and rename

• The chat view now supports switching between conversations via a dropdown selector.
• A future change could add the ability to delete or rename conversations from the selector.

Image paste support in chat

• The chat view now shows a notice when images are pasted. A future change could save pasted images to the vault and attach them as multimodal context.

Dashboard layout customization in the onboarding interview

• The dashboard layout is manually configurable via the "Customize dashboard" button in the dashboard header, with 11 widgets supporting reorder, show/hide, collapse/expand, and size (compact/standard/expanded).
• A future enhancement could add a dashboard layout phase to the onboarding interview (e.g. extending the "style" phase, or adding a new phase after "confirmation").
• This would require:
  1. Adding dashboardLayout to the OnboardingConfig type in src/onboarding/OnboardingTypes.ts
  2. Updating the interview system prompt in src/onboarding/InterviewEngine.ts to ask about widget preferences
  3. Applying the interview-derived layout after interview completion in src/main.ts

The OnboardingConfig already has a style.dailyNoteLayout field, so a style.dashboardLayout field would follow the same pattern.

Dashboard telemetry health summary

• The dashboard now shows provider icons with green/red readiness dots. A future change could show live health check status (connected/error) next to each icon.

Voice Prompt Modal — candidate-less mic guard

• The voice prompt modal now shows a warning when no transcription providers are configured. A future change could disable the mic button entirely, matching the chat view behavior.

Quality, security, and release controls

The test suite currently contains 342 tests:

• 167 core tests — build integrity, parsers, byte-safe RPC framing, subprocess integration, task queue, recovery, provider fallback, capability registry, user memory, system prompts, project manager, composer fuzzy matching, @-mention engine, fallback pipeline, and STT/TTS adapters
• 153 ReAct and subsystem tests — roles, evaluation, traces, workflows, Bases, chat context, action cards, audio, JIT lifecycle, RAG, memory, CLI, locks, and stress scenarios
• 22 provider tests — XAI provider, OpenRouter model metadata, transcription candidates, and model matrix integration

CI runs on Windows, macOS, and Linux across Node 20, 22, and 24 with:

1. Repository sanitization
2. Typecheck
3. Zero-warning lint
4. Full tests
5. Standardized benchmarks
6. 25% performance-regression gate
7. Production package validation
8. Clean release-surface verification

Release automation repeats the validation, builds a clean three-file plugin package, attests artifact provenance with Sigstore attestations, verifies each published asset cryptographically, and creates a GitHub release. The package metadata and manifest version are currently 1.7.2, with Obsidian 1.13.0 as the minimum supported app version.

The local community-plugin validator currently passes with 0 errors.

Troubleshooting

Pi cannot be found

Run:

pi --version
npm install -g @earendil-works/pi-coding-agent

Then use Settings → Command Center → Core → Pi harness path and click 🔍 Detect (or type a custom path). Detection is non-blocking; the button shows ⏳ Detecting… while it runs and reports ⚠️ Not found if the binary is still missing. Missing-binary errors fail fast rather than entering a retry loop.

A local provider is unavailable

• Confirm Ollama or LM Studio is running.
• Verify the base URL in the provider card.
• Click Refresh Models, then run the provider health check.
• Ensure the selected route references a model reported by the local server.

Command Center says the vault is uninitialized

Run Command Center: Start Setup / Onboarding Interview and complete confirmation/synthesis. Both .command-center/config.json and .command-center/style-guide.md must validate before operational services start.

A workflow does not run

Check that:

• The active file is Markdown or Canvas and contains workflow metadata.
• Step IDs are unique.
• Every dependency exists and the graph is acyclic.
• Required inputs are supplied.
• Provider/Pi routes are enabled and healthy.
• A Base file references a valid workflow path.

A destructive tool is paused

Open the Command Center dashboard, inspect the Mutation approvals card's target list and diff preview, then choose Approve & apply or Reject. Closing the dashboard rejects pending confirmations safely.

“All transcription providers failed … model does not exist”

STT model IDs are per-provider. The error The model 'openai/gpt-4o-mini-transcribe' does not exist means a slug meant for one provider (here, OpenRouter routing) was sent to another (e.g. xAI). Fix it in Settings → Accessibility & Speech → Per-provider speech-to-text models: set the slug each provider actually accepts, or leave the field blank to use that provider's built-in default (grok-stt for xAI, whisper-1 for OpenAI, etc.). The global Speech-to-text model field is no longer broadcast to every provider — it only applies to providers that have no built-in default.

“All transcription providers failed … no speech detected”

The provider processed the audio successfully but returned an empty transcript (near-silent input, background noise, or a Whisper silence-hallucination artifact that was stripped). Speak closer to the microphone, or switch the Speech-to-text provider to one with a higher-quality STT model. Recording shorter than 500 ms is treated as an accidental mic tap and never sent.

Future implementation

The following features have infrastructure stubs and are ready for implementation when needed:

Video generation (OpenRouter)

Command Center includes static URL builders for OpenRouter's video generation API:

• POST /api/v1/videos — submit a video generation request
• GET /api/v1/videos/{id} — poll generation status
• GET /api/v1/videos/{id}/download — download generated video
• Models are discovered dynamically through /api/v1/models/user when they become available
• Vault media ingestion stubs (video MIME types, isVideoFile()) are ready in image-utils.ts

Image generation (OpenRouter)

Image generation models are registered in the OpenRouter provider (openai/gpt-5-image, google/gemini-3.1-flash-image, etc.) and the getImageGenerationUrl() helper returns POST /api/v1/images/generations.

xAI Realtime API

xAI's realtime API (GET /v1/realtime WebSocket) supports voice-in/voice-out and function calling. The XAIProvider class (src/providers/xai.ts) implements chat completions plus native STT (POST /v1/stt) and TTS (POST /v1/tts, GET /v1/tts/voices). TTS is now wired into the spoken-output pipeline via TtsAdapter (see Text-to-speech above); the WebSocket realtime transport remains a future addition.

OpenRouter Responses API

The newer POST /api/v1/responses format is a future migration target alongside the existing chat completions endpoint. Adopting it would enable streaming reasoning tokens, server-side tools (web search, code interpreter, MCP), and compaction; no Responses endpoint is wired up today.

Community and support

• Read CONTRIBUTING.md before proposing changes or opening a pull request.
• Follow the Code of Conduct in all project spaces.
• Use GitHub Discussions for questions and community support; see SUPPORT.md.
• Report vulnerabilities privately according to SECURITY.md—never in a public issue.

Donations and developer support

Command Center is free and MIT-licensed. If it improves your workflow and you would like to thank the developer, you can make an optional donation:

☕ Buy Dustin a coffee

The same link is available as a branded Buy Me a Coffee button in the support card at the bottom of Settings → Command Center. Donations are optional, do not unlock features, and are not required for support or updates.

License and attribution

Command Center is released under the MIT License.

Third-party projects remain under their respective licenses. See ATTRIBUTIONS.md for the audited dependency inventory and distribution boundaries, including:

• Obsidian API — MIT; host API/types, not bundled as Obsidian itself
• esbuild — MIT; build tooling
• Pi coding agent — MIT; optional external RPC companion
• TypeScript, ESLint, CodeMirror/Lezer, Node.js, Electron, and other development/host components

The lockfile audit found no GPL-family, SSPL, BUSL, or undeclared package licenses. Repository sanitization, credential-boundary checks, clean-room packaging, and provenance attestation are part of the publication workflow.

“Obsidian” is a trademark of Dynalist Inc. Command Center is an independent community plugin and is not endorsed by Dynalist Inc. or the Pi maintainers.