Skip to content

Milestones

The v1 delivery plan from spec 16.

v1 — the minimum viable product

#DeliverableStatus
M0Monorepo scaffold — apps/{cli,llama,dashboard}, root Taskfile, go.work, .gitmodules, AGENTS.md per app✓ shipped (f0e2755)
M1apps/llama cgo binding to llama.cpp (submodule init + libllama.a + binding tests on linux + darwin + windows)🚧 pending
M2apps/cli skeleton: cobra root, init, doctor. init scaffolds + top-N model picker; doctor reports model presence + toolchain✓ shipped (38d6c46)
M3wiki generate v1: data extraction + commands exec + wikigen orchestrator wired to the local backend; honors wiki-instructions.md + schema.md; frozen: true opt-out🚧 pending
M4Wiki shape: typed articles, frontmatter contract, wiki/index.md, wiki list / show / edit🚧 pending
M5build v1: skills validation + RAG chunk index + manifest assembly + zip output; --self-contained and --to s3://; demo-question smoke test🚧 pending
M6Chat tool calling: search_wiki, read_wiki, list_wiki, search_rag (in-process; apps/llama exposes a function-calling-aware Chat)🚧 pending
M7Skills: SKILL.md + four-stage progressive disclosure (no CLI subcommand; folder-driven); optional runtime: field checked at build + doctor🚧 pending
M8serve with /api/* + embedded dashboard at /; dashboard surfaces demo: as suggested prompts🚧 pending
M9import (default mode): zip unpack, manifest validation, https + file source🚧 pending
M10import with --self-contained zips: model bundling round-trip🚧 pending
M11local-agents welcome screen; end-of-command "next:" suggestionspartial — welcome screen ✓ via M2; per-command "next:" suggestions are per-command work
M12Native CI release builds per target (linux + darwin + windows)🚧 pending
M13Bundle compat with offline-llm-knowledge-system verified both directions🚧 pending
M14eval v1: pluggable judge backends (claude / openai / command); --generate synthesizes a test set from data/; multi-turn scenarios; pass-threshold exit code🚧 pending

v2 — the quality tier

  • Pluggable wikigen backends: claude-code, codex, generic command. The v1 wiki.generator config slot already exists.
  • OpenAI-wire-compatible /v1/* surface so external SDKs work unchanged.
  • OCR for images in data/ (Tesseract offline + vision LLM online).
  • More PDF extractors (poppler fallback for hard cases).
  • windows/arm64 build.

v3 — the integrations tier

  • gs://, azure:// sources for import + build --to.
  • wiki edit opens a real markdown editor with live preview.
  • Build cache hit visualization in the TUI.
  • Bundled extractors for common formats (DOCX, XLSX, EPUB).

pocket llm — local-first, offline, no telemetry. MIT licensed.