plantool

LLM-led project-planning tool: an MCP server whose tools interview the user, record every planning fact as a typed row in SQLite, and refuse to advance past mechanical gates. Spec: plantool_stage1_spec.md (rev 2 — source of truth).

Layout

• engine/ — plain Python, no LLM calls: schema, DB wrapper, validation, next_gap() walker.
• server/ — FastMCP stdio server + interview scripts (delivered inside tool results).
• workspace-template/ — copy this directory to start planning a project.
• tests/ — pytest suite (fixture DBs).

Setup

py -3.12 -m venv .venv
.venv\Scripts\python -m pip install mcp pyyaml pytest

Use

Copy workspace-template/ to a new folder, open your MCP-capable CLI (Claude Code first) there, and talk. The .mcp.json launches the server over stdio; plan_status() bootstraps the model.

Tests

.venv\Scripts\python -m pytest

Build status

Sessions A–D of 5 (spec §9) are in:

• A — schema v1, plan_start / plan_status / next_gap, first submit tools.
• B — the full submit surface (use cases, CRUD, state machines, components, contracts, dependencies, failure modes, decisions, questions, conflicts).
• C — enforcement: gates 1–6, spikes, engine conflict sweeps, plan.yaml export/reimport.
• D — the dogfood harvest: supersede/retire/confirm correction path with lineage (claim rows are never mutated or duplicated), dangling-ref checking on every batch submit, plan_status digest + get_rows targeted reads, dismiss_gap, and the interview-conduct rewrites (divergence rounds, recorded challenges, file-question-first).

Session E (red-team script, get_plan_pack, stages 7–8 gates, export_plan, freeze_plan) closes stage 1.