feat: LLM assistant integration for natural language task capture and scheduling suggestions #2

New issue

Open

opened 2026-07-05 11:02:34 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-07-05 11:02:34 -07:00

Collaborator

Context

The Focus Flow design doc (Unified Product Design Summary) lists "intelligent
assistant features" as a wishlist item. CF's circuitforge-core LLM module
provides a local-first LLM layer that aligns with Focus Flow's privacy principles.

Proposed scope

Natural language task capture

"Add a dentist appointment Thursday at 2pm" → parsed into a structured task with
inferred type (inflexible), time, duration estimate.

Uses a small local model (llama3.2:1b, mistral:7b) — no cloud required, no data
sent anywhere. Runs via CF's existing ollama/vllm integration.

Scheduling suggestions

When the day is over-filled or a task has been pushed repeatedly, offer plain-language
suggestions: "You've pushed 'Write report' 3 days in a row. Split it into smaller
steps?" — framed as observations, not nagging.

End-of-day summary (V2.0 weekly reports)

LLM-generated plain-language summary of what was completed, what was pushed, and
patterns. Opt-in, local-only.

Architecture

LLM calls go through a thin adapter in packages/scheduler_core or a new
packages/scheduler_llm package
Adapter interface: LlmProvider (local vs cloud-optional) — CF provides the
implementations
UI triggers (compose bar, push button long-press, end-of-day screen) remain
in the Flutter layer

Privacy note

Local inference is the default and the required free-tier path. Any cloud LLM
option is opt-in, gated, and uses CF's privacy architecture (no PII logging,
no behavioral profiling).

License note

LLM integration code contributed by CF will be AGPL-3.0 consistent with the rest
of Focus Flow. CF's cloud infrastructure layer (BSL 1.1) operates separately and
is not bundled into this repo.

## Context The Focus Flow design doc (Unified Product Design Summary) lists "intelligent assistant features" as a wishlist item. CF's `circuitforge-core` LLM module provides a local-first LLM layer that aligns with Focus Flow's privacy principles. ## Proposed scope ### Natural language task capture "Add a dentist appointment Thursday at 2pm" → parsed into a structured task with inferred type (inflexible), time, duration estimate. Uses a small local model (llama3.2:1b, mistral:7b) — no cloud required, no data sent anywhere. Runs via CF's existing ollama/vllm integration. ### Scheduling suggestions When the day is over-filled or a task has been pushed repeatedly, offer plain-language suggestions: "You've pushed 'Write report' 3 days in a row. Split it into smaller steps?" — framed as observations, not nagging. ### End-of-day summary (V2.0 weekly reports) LLM-generated plain-language summary of what was completed, what was pushed, and patterns. Opt-in, local-only. ## Architecture - LLM calls go through a thin adapter in `packages/scheduler_core` or a new `packages/scheduler_llm` package - Adapter interface: `LlmProvider` (local vs cloud-optional) — CF provides the implementations - UI triggers (compose bar, push button long-press, end-of-day screen) remain in the Flutter layer ## Privacy note Local inference is the default and the required free-tier path. Any cloud LLM option is opt-in, gated, and uses CF's privacy architecture (no PII logging, no behavioral profiling). ## License note LLM integration code contributed by CF will be AGPL-3.0 consistent with the rest of Focus Flow. CF's cloud infrastructure layer (BSL 1.1) operates separately and is not bundled into this repo.