jmatzkin1980/ignite-sentinel

> A repo-local, local-first system that matures raw client input into traceable, agent-ready requirements — governed from first note to backlog.
Most tools generate documents. Ignite Sentinel does something harder: it matures requirements. You feed it the messy first version of a client request — a note, a screenshot, a half-formed idea — and it works the material through a governed lifecycle until it becomes a project brief, PRD, specs, backlog, quality artifacts, and a full traceability graph. Nothing is invented: every claim is backed by citable evidence, and everything missing stays an explicit GAP- or [PENDING INPUT].
The source of truth is always the versionable files under workspaces/PROJECT_ID/. Local memory is a retrieval aid, never the authority. No client content leaves your machine.
text
raw input ──/ingest──▶ discovery gaps ──(optional)──▶ /annotate · /challenge · /scrutinize · /assume
│ (agentic deepening)
/resolve-gaps ──▶ /maturity ──▶ /brief
(closes confirmed, (gates, (evidence-compiled,
normalizes EARS) telemetry) per-section readiness)
/specs ──▶ /backlog ──▶ /quality ──▶ /trace · /health · /validate
## Contents
Why it's different
Who it's for
The lifecycle
Commands
Quick Start
Driving it from chat
Governed backlog
Resolving change over time
Command protocol
Local memory
Workspace layout
Design rules
Portability
Verification
FAQ
Documentation
## Why it's different
Purpose-built for maturing requirements — not a generic spec generator. Unlike general spec-driven (SDD) or BMAD-style frameworks, Ignite targets the specific BA/Product pain: working raw client input into a robust spec. Discovery is the heart of the system, not an afterthought.
It matures, it doesn't fabricate. Every downstream artifact traces back to confirmed evidence. What isn't known yet stays a visible GAP- or [PENDING INPUT] — the framework refuses to invent scope.
The lifecycle is governed, not just generated. Gates stop you from building specs on an immature requirement or a backlog on stale context. When a command blocks, it tells you the correct previous step.
The agent is a sanctioned analyst, not a free hand. Agents can deepen discovery (/annotate, /challenge, /scrutinize), register BA-owned assumptions (/assume), enrich the PRD (/compose) and refine the backlog (/refine-backlog) — but the runtime validates every contribution against a verbatim quote before it touches an artifact. You stay in control.
Vector-backed progressive disclosure. A local vector index (LanceDB, or a deterministic fallback) feeds each phase and activity only the context it needs — generation reads the exact section instead of rereading the whole workspace. Retrieval is contextual (situational prefixes per chunk), re-scored by recency and domain coverage, and can be queried as an episodic timeline ("what changed and when") or biased toward recent context. Discovery and /sync emit pointer-only focus packs, multi-section packs enforce a global token budget and de-duplicate evidence, and a soft "needs-context" gate in /health nudges flows to consult focus instead of reading everything — all degrading identically without LanceDB.
Local-first by default. No remote MCP, external vector database, or external embedding service for client content. Runs fully on a locked-down VDI; LanceDB is optional and degrades to deterministic json-hybrid; semantic embeddings are optional local packages only.