Skip to content

Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS

12.4 KB
Raw Blame

Chief of Staff Enhancement (M35)

Status: This M35 milestone spec shipped, except the Phase 5 agent-architecture services (agentGateway.js, errorRecovery.js, agentRunCache.js, contextUpgrader.js) and sessionDelta.js, which were never built under those names. Their rows below are marked "(not built)" and kept for historical legibility.

Comprehensive upgrade from reactive task executor to proactive autonomous agent with hybrid memory, missions, and local model integration.

Architecture

New Services

Service Purpose
server/lib/bm25.js BM25 algorithm with IDF weighting
server/services/memoryBM25.js BM25 index manager for memory search
server/services/sessionDelta.js Session delta tracking for pending bytes/messages (not built)
server/services/toolStateMachine.js Tool execution state machine
server/services/agentGateway.js Request deduplication and caching (not built)
server/services/errorRecovery.js Error analysis and recovery strategies (not built)
server/services/agentRunCache.js Agent output caching with TTL (not built)
server/services/eventScheduler.js Cron-based event scheduling
server/services/executionLanes.js Concurrent execution lane management
server/services/missions.js Long-term goal and mission management (removed, #6826 — nothing in production ever created a mission; superseded by the Goals system)
server/services/lmStudioManager.js LM Studio model discovery and health
server/services/localThinking.js Local model completions
server/services/thinkingLevels.js Dynamic model selection (removed, #8149 — tier/model choice is explicit; see MODEL_TIERS.md)
server/services/contextUpgrader.js Complexity analysis for model upgrade (not built)

Features

Phase 1: Hybrid Memory Search

  • BM25 algorithm with IDF weighting and inverted index
  • Reciprocal Rank Fusion (RRF) combining BM25 + vector search
  • 40% BM25 / 60% vector weighting for optimal retrieval
  • Session delta tracking for pending bytes/messages

Phase 2: Proactive Execution

  • Event scheduler with cron expressions and timeout-safe timers (clamps to 2^31-1)
  • Execution lanes: critical (1), standard (2), background (3) concurrent slots
  • Mission system for long-term goals with sub-tasks (removed, #6826)
  • Mission-driven task generation in evaluation loop (removed, #6826)

Phase 3: Local Model Integration

  • LM Studio availability checking and model discovery
  • Quick completions for local thinking without cloud costs
  • Memory classification using local models
  • Embeddings via local LM Studio

Phase 4: Dynamic Model Selection (removed)

Priority/description-driven thinking levels and automatic tier changes were removed in #8149: a task runs on the model or tier it names, else the provider default. See MODEL_TIERS.md.

Phase 5: Agent Architecture (not built)

Phase 5 services were never built under these names; see the Architecture section above. What was actually shipped for agent failure handling is described in Error Handling.

  • Tool execution state machine (IDLE → START → RUNNING → UPDATE → END → ERROR) — not built
  • Agent gateway with request deduplication and 10-minute cache — not built
  • Error recovery with 6 strategies: retry, escalate, fallback, decompose, defer, investigate — not built (see Error Handling)
  • Agent run cache for outputs, tool results, and contexts — not built

Execution Lanes

Lane Concurrent Slots Purpose
critical 1 Emergency fixes, blocking issues
standard 2 Normal development tasks
background 3 Self-improvement, documentation

Error Recovery Strategies

Strategy When Used
retry Transient errors (rate limit, timeout)
escalate Persistent failures requiring human review
fallback Model unavailable, try alternative
decompose Complex task, break into subtasks
defer Non-urgent, schedule for later
investigate Unknown error, create investigation task

API Endpoints

Route Description
GET /api/lmstudio/status Check LM Studio availability
GET /api/lmstudio/models List loaded models
POST /api/lmstudio/completion Local model completion
POST /api/lmstudio/analyze-task Analyze task complexity
POST /api/lmstudio/classify-memory Classify memory content

Design Decisions

  1. Mission Autonomy: Full autonomy - COS can implement changes, run tests, commit without approval for managed apps
  2. Model Usage: Local-first with LM Studio, no cloud API costs for thinking
  3. Persistent Mind model control: The default profile is human-owned. An optional grant permits one approved local preset on a self-directed wake; the mind cannot rewrite provider configuration.
  4. Failure Investigation: Unattended by default - an [Auto] Investigate agent failure task runs without approval, because an isolated agent failure is exactly the work CoS exists to diagnose for itself. Approval is reserved for a failure loop, where another unattended agent would only repeat what already didn't work: the same cause investigated again within 24h (repeat-fingerprint), or the circuit-breaker hour already one slot from its cap (failure-storm). The reason is stamped on the task as the producer-agnostic metadata.approvalReason (investigation-loop:<signal>), which the Tasks UI turns into a hint on the APPROVE button. See resolveInvestigationApproval in server/services/agentErrorAnalysis.js.

Persistent Mind task filing

The Persistent Mind has one explicit, default-off typed action, cos.create-task. When the grant is enabled, each wake receives a bounded server-built catalog of runnable apps and enabled CLI/TUI coding providers. The catalog includes each provider's selectable models and effort levels, and marks apps whose effective work tracker supports the issue-only Plan & File Issue workflow.

The mind may select the provider, model, effort, priority, delivery mode, and PR completion policy per request. Implementation mode requires one of review-then-merge, merge-on-green, or leave-open; plan-only mode is only valid for GitHub/GitLab apps and is normalized to the existing unattended plan-task --yes no-code/no-PR contract. The queue path revalidates every choice after inference, so stale prompts, disabled providers, invented model ids, and unsupported trackers fail closed.

Task model selection can be restricted in Persistent Mind Tools with an exact provider/model allowlist. An empty list preserves the unrestricted catalog; a non-empty list is applied to both the model choices sent to the mind and the server-side queue admission check, which supports subscription-specific or local-only task lanes. A removed or malformed allowlist fails closed.

GET /api/cos/mind/tools returns the authority inventory and the managed-app roster. It does not list coding providers, models, or effort levels — the mind still receives that redacted catalog in its wake-time task prompt, and Mind Tools restricts models only through the task-model allowlist. Broader PortOS APIs remain governed by their own authenticated route contracts and are not silently made callable by enabling task filing.

Persistent Mind phone calls

voice.call-user is the mind's only outward-reaching authority, and like every other grant it is off by default — an install upgrading with the mind already running gains nothing until the user turns it on in Mind → Tools. It lets the mind place a FaceTime Audio call to the single handle configured in Settings → Voice → FaceTime Audio. The request the model returns (callRequest: { reason, openingLine }) carries no recipient, so a confused or compromised turn cannot dial anyone else.

The gate lives in server/services/persistentMindCallCapability.js and runs after inference. A call is placed only when all of these hold:

Check Suppression reason
The grant is on (re-read from stored config) not-granted
The optional FaceTime feature is enabled feature-disabled
Voice is enabled and an identity is saved voice-disabled, identity-unconfigured
No browser tab can speak the message instead tab-available
The user's local time is outside voice quiet hours quiet-hours
No call is already up call-in-progress
≥30 minutes since the last call, ≤3 calls in the last 24 h too-soon, rate-capped
A call-host tab is attached and the helper dials no-call-host, dial-failed

The caps are counted in durable Persistent Mind state (callHistory), so a restart, a crash, or a supervisor rewire cannot hand back a fresh allowance. The budget is spent only on a call that actually went out. Every decision — placed or suppressed, and why — is appended to the trajectory as mind.call.requested plus mind.call.placed / mind.call.suppressed; the dialed handle is never written there. When a request is refused, the turn's reply is appended with the reason, so the mind cannot leave the user waiting for a phone that never rings.

A placed call runs with the mind's persona and a bounded briefing of its trajectory, speaks the openingLine the moment the far end picks up, and hands the transcript back as a Persistent Mind message on hangup, so the next wake knows what was said.

Critical-notification escalation (Settings → Voice) uses the same gate and the same budget, but is authorized by its own facetime.escalateCritical setting rather than by the mind's grant. When a critical notification is still unread after facetime.escalateAfterMinutes (default 10) and no voice tab is available, PortOS asks to ring the user with the notification as the opening line. Off by default.

Test Coverage

Test File Module
server/lib/bm25.test.js BM25 Algorithm
server/services/toolStateMachine.test.js Tool State Machine
server/services/executionLanes.test.js Execution Lanes
server/services/errorRecovery.test.js Error Recovery (not built)
server/services/agentRunCache.test.js Agent Run Cache (not built)
server/services/missions.test.js Missions Service (removed, #6826)

Related Features

Self-directed local thinking presets

In the Mind settings drawer, Models → Self-directed local thinking lets the user enable chooseThinkingPreset and approve exact local presets. It defaults to off on new and upgraded installs and is independent of task-agent model permissions. Saving preferences and loading the catalog perform no inference.

Only configured direct local API runtimes qualify. Account-backed routes, CLI/TUI wrappers, and unknown routes are excluded. Approval binds the preset, provider name, endpoint and runtime; editing these requires renewed approval. Removing a model, preset, or supported effort refuses the pending selection.

The semantic actions mind.thinking-presets and mind.request-thinking-preset expose the approved catalog and a bounded { presetId, reason } request. One request borrows a preset for the next self-directed wake; queued human messages and the current turn keep their own routes. The default profile, memories, and identity are unchanged.

Only one request may be pending or active. At most three are accepted per rolling 24 hours, with a 30-minute gap. Admission also advances the recorded attempt time so a delayed wake cannot evade these limits. Cancellation, failed admission, interruption and restart do not refund allowance. Failed attempts retire the alternate and schedule an ordinary wake on the default profile. Existing endpoint slots, CoS capacity, autonomy and per-call budgets still gate every execution. No tool can extend its own session or edit these permissions.

The same Models panel cancels pending requests and shows recent concise reasons and outcomes. Disabling the grant cancels the pending request; revoking it during execution refuses further provider calls. Execution receipts in Sessions show the actual route and outcome without exposing hidden reasoning. Durable request records live alongside the existing machine-local mind state; migration 348 preserves prior profiles, grants, and counters.