Skip to content

roadmap.md

Latest commit

 

History

History
95 lines (82 loc) · 5.06 KB

roadmap.md

File metadata and controls

95 lines (82 loc) · 5.06 KB
Raw

Roadmap

Status

v0.1 is feature-complete against MY-DAEMON-SCAFFOLD.md's Phase 0–3 spec, the M1–M4 adaptive-memory milestones, and the project-level acceptance criteria. The end-to-end pipeline (ingest → retrieve → synthesize → log) works against a real Obsidian vault, and the daemon envelope around it — config resolution, atomic state writes, preflight checks, backup/restore, supervisor, scheduling — is in place. The three background-agent writeback jobs ship but are gated behind agent.enabled: false until you confirm their thresholds with --dry-run.

Done

  • Phase 0 — Skeleton. Repo, pyproject.toml, uv setup, ruff + pytest, config.py, models.py, sample test vault, docker-compose.yaml for Qdrant.
  • Phase 1 — Ingest pipeline. Vault reader / parser / chunker with tests; sentence-transformers embedder; Qdrant + NetworkX stores with persistence; daemon ingest and daemon status.
  • Phase 2 — Retrieval + query. Seed / expand / orchestrator, Anthropic client, synthesis prompt, daemon query, SQLite feedback log.
  • Phase 3 — Incremental + polish. Manifest-based incremental ingest, daemon graph stats (PageRank + tag degree), README, end-to-end test.
  • Hybrid retrieval (2026-05-16). Sparse BM42 vectors stored alongside dense, fused server-side in Qdrant via RRF. Toggle with embeddings.hybrid in config.yaml.
  • Background-agent writeback (2026-05-16). Three new commands: daemon extract, daemon link, daemon reflect. Shared safety discipline in vault/writer.py. Defaults to Haiku 4.5 via llm.batch_model. Tkinter setup window with a Windows Task Scheduler checkbox for daily reflection.
  • NiceGUI chat (daemon chat). Streaming responses, brand-aligned dark theme, lazy retrieval pipeline.
  • Setup window (daemon setup). Vault picker + API-key persistence.
  • Adaptive memory loop, M1–M4 (2026-05-17). Adaptive edge weighting from feedback signals; the snapshot mechanism; structural-pattern analysis over frozen snapshots (Louvain communities, edge betweenness); and the observer LLM that reads those reports and writes the user-facing letter.
  • Daemon envelope (2026-07-26). Config search order with loud missing-config failure and config-dir-anchored paths, atomic graph and manifest writes with an inter-process lock, embedded Qdrant as the default, daemon doctor, backup/restore, and the daemon run supervisor plus daemon schedule.
  • API key out of plaintext (2026-07-27). The key resolves environment → OS credential store → legacy .env, and daemon setup writes only to the credential store, migrating and stripping any plaintext copy it finds. See Getting Started.

Next

In priority order:

  1. Multi-embedding spaces. A second embedding alongside the semantic one — emotional valence, entity-based, code-vs-prose. Hybrid retrieval already proves the multi-signal plumbing; this just adds another slot.
  2. Obsidian plugin / file watcher. Live ingest as you save. daemon run already watches the vault; the remaining work is the editor-side plugin.
  3. Local LLM for synthesis. For users who want fully-offline operation end-to-end. The streaming surface already abstracts the provider; a local-model client implementing synthesize / synthesize_stream slots in.

The open design question behind all of it: seeds dominate the candidate pool and seed-picks currently reinforce nothing, so the learning loop closes mechanically but has little to learn from. Deciding what "picking" should mean comes before scaling any more learning machinery — see the AI Suggestions section of PROJECT_MANAGEMENT.md.

Deferred indefinitely (for a reason)

  • Heavy auth / multi-user. The daemon is intentionally a single-person tool that runs on your machine against your vault.
  • A SaaS frontend. Same reason. The brand says "companion," not "service."
  • Vault rewriting beyond ## Agent Notes and tags. The agent edits a bounded surface and snapshots before every change. Rewriting prose itself would cross the line from companion to ghostwriter.

Open agent-suggestions (deliberately not implemented)

From PROJECT_MANAGEMENT.md:

  • daemon select <feedback_id> <#> for candidate-selection feedback. One CLI flag away. Held back to nail down the signal shape first.
  • daemon query "..." --skip <feedback_id> for "more like these." A second-three-results request. Useful in CLI but the eventual home is the GUI; may not be worth investing twice.
  • daemon graph todos from dangling wikilinks. The graph already keeps dangling targets as placeholder nodes. A command that surfaces them is "notes you mean to write" — a feature, not a bug.
  • Seed the daemon's own vault with a "purpose" note. So when the daemon ingests its own vault, it can answer "why am I building this" in the user's voice.