Skip to content

MVP_AND_ROADMAP.md

Latest commit

 

History

History
78 lines (58 loc) · 8.18 KB

MVP_AND_ROADMAP.md

File metadata and controls

78 lines (58 loc) · 8.18 KB
Raw

MVP plan, current status, gaps, and suggestions

This document captures the product vision and MVP we aligned on in planning, what exists in this repository today, what is still missing for a full MVP or 1.0, and suggestions for the next iteration.

It is a living document—update it when major features land.

1. The north star

Open Code aims to be a logic-first, agentic coding workspace: users shape the project as readable cards/topics first, then parallel agents implement approved cards in isolated git branches/worktrees.

  • Logic board + agent runtime + memory as infrastructure (not a chat bolt-on).
  • Card-per-agent isolation: each runnable logic card gets its own branch/worktree; code conflicts go to merge agents, logic conflicts go to humans.
  • Readable source of truth: structured JSON plus generated Markdown under logic/ so project logic can be reviewed on GitHub.
  • Default open model: Gemma 3 4B (pinned) on localhost, with Auto model selection and everything packaged so users are not required to install Ollama or other runtimes by hand in a future consumer build (Ollama-style API remains a compatibility target).
  • Autodrive / Autopilot: a multi-turn conversational agent (like a persistent chat) with pause/stop, scope, and review/undo where the product contract requires it.
  • Multi-mindset self-critique (QA, tests, scalability, fault tolerance, extensible) with per-project toggles (e.g. turn off “scalability” for a tiny project so the “brain” does not think about it).
  • Discriminator (persistent): run enabled mindsets on prompts, proposed edits, and commits (or equivalent), and log results into memory.
  • Two-layer memory:
    • Raw evidence: messages, files, diffs, tool output, user actions, and later media (screenshots, audio/video) with user consent and retention policy.
    • Semantic layer: summaries, facts/triples, preferences, goals, decisions, timelines, open tasks (lightweight in MVP, richer over time). Links back to raw ids for every surfaced claim.
  • Rust for fast subsystems, especially the memory pipeline; TypeScript for VS Code / workbench integration and most UI. Re-use Code-OSS/VS Code built-ins (theming, diff, commands, webviews) before building bespoke UI.

2. What is implemented in this repo (bootstrap)

Area What exists
Logic workspace crates/open-code-logic: topics/cards schema, branch/worktree planning, conflict classification, generated documentation, and tests for the multi-agent model.
Desktop app scaffold packages/desktop: Tauri + Svelte UI for topic navigation, card inspection, conflict review, running agent badges, and VS Code handoff.
Readable project state logic/open-code.project.json, logic/open-code.paper.md, and presentation docs: GitHub-readable starter state for the new product model.
VS Code extension packages/extension: commands (Autodrive webview, run agent, review diff, health, clear/show memory, discriminator on commit text in active editor), local Gemma/Ollama-style HTTP provider, Open Code Dark theme, inline completion hook (uses local model when available).
Agent loop Proposes changes via local LLM, shows native diff editor, computes line hunks, supports Accept All / Choose Hunks / Reject; memory logging of diffs, hunk ids, and actions. Undo = editor undo.
Memory Rust open-code-memory + open-code-memoryd: SQLite raw + semantic tables, HTTP API on 127.0.0.1, append/clear/recent; hashed payload field on raw.
Discriminator (partial) Mindset toggles in settings; onPrompt / proposed edits / commit path hook; writes raw + semantic (critic summary) when the LLM responds. Not full pluggable critic modules or gating of auto-apply.
Build / CI npm workspace, Cargo.lock, GitHub CI matrix on Ubuntu/macOS (fmt, clippy, test, build, bootstrap checks), fork-health for lineage checks.
Docs BUILD, GOVERNANCE, RELEASE, SECURITY, FORK_AND_RELEASE

3. What is missing (relative to the full plan MVP / 1.0)

Gap Notes
Code-OSS / Electron product fork The extension is not yet merged into a full editor fork. That is a large follow-on for signed desktop apps.
Packaged local inference (no extra installs) No bundled Ollama/llama.cpp sidecar in the extension VSIX; users must run a server at openCode.gemmaBaseUrl (e.g. Ollama) or use offline stubs. A one-installer story is TBD.
Gemma “pinned in product” Default URL assumes a local server; extension setting now defaults to gemma3:4b, but release builds still need exact artifacts and checksums in a release manifest.
Line-by-line / logic-by-logic review Basic line-hunk selection exists; still missing a dedicated review tree, inline decorations, stale-document detection, and AST-aware acceptance.
Full raw evidence types Multimedia (screens, audio, video) and rich tool output capture are not fully wired (schema can grow).
Semantic layer depth No automatic triple extraction, nightly compaction, or vector/ANN retrieval; summaries are only where the discriminator and flows write them.
Memory scopes Host / user-wide stores and credential_ref audit are not productized; per-project is approximated via workspace folder id.
Autodrive Webview + memory is present; no formal step budgets, allowlists, or block of auto-apply.
Mindset pluggins One LLM call with a list of enabled mindsets, not separate critic modules, budgets, or async critic queues.
E2E tests npm run test:e2e now covers docs/bootstrap, desktop preview workflows, extension activation, Rust logic/memory, workspace build, and a strict local Ollama/Gemma response. CI runs npm run test:e2e:ci, skipping only the multi-GB model call.

4. Suggestions (next steps, prioritized)

  1. Inference packaging — SPIKE: bundle Ollama-compatible binary or llama.cpp sidecar (Rust is a good fit) per OS; one managed first-run or embedded download for Gemma 3 / Gemma 3n weights. Keep the app talking only to a small HTTP API on localhost.
  2. Memory: profile then optimize — Keep Rust for ingest/index hot paths; add migrations and version table. Add retention and export for GDPR-style deletion.
  3. Review UX — Move from “replace whole file” to patch hunks using workspace edits + a dedicated review tree or inline decorations consistent with README must-haves.
  4. Vector retrieval — Add optional embedder and ANN (or SQLite extension) for semantic search; always point back to raw rows.
  5. Credentials — Wire SecretStorage for any non-local API keys; keep memory as metadata only (SECURITY).
  6. Promote to fork — When stable, place the extension under extensions/ in a Code-OSS fork and wire product.json for a branded app (GOVERNANCE, FORK_AND_RELEASE).
  7. Test matrix — Keep extending the E2E matrix toward packaged-app flows, managed runtime download, memory retention/export, and fork-health checks.

5. How this relates to the original README

The Must-Have UX and Memory Scopes in the README are still the contract. This file explains how far the bootstrap has progressed and what remains before claiming MVP or 1.0 against that contract.

Definition of a stronger MVP candidate (internal bar): packaged or documented local Gemma 3 path, stable two-layer memory with revoke/clear, autodrive with clear stop semantics, and at least two real mindset + documented discriminator hooks on prompt + edit + commit—all without requiring undisclosed plaintext secrets in SQLite.

For release mechanics, see RELEASE and FORK_AND_RELEASE (“Definition of Releasable”).