Glossary

Shared vocabulary for ContextMint product, support, and documentation.

ContextMint Engine

Desktop tray application that runs the Python API server, manages Ollama, indexing, logs, and operator UI tabs.

ContextMint (extension)

VS Code extension — chat webview, Context Lens, packs panel, quality/governance UI. Connects to Engine over HTTP.

Workspace ID

12-character hex identifier for a registered folder. Stable per repo path; used in API calls and audit.

Context lane

Repo (full RAG), Work (packs/attachments only), or Hybrid (RAG + recommendations). Selected in compose ⋯ → Context.

Progressive RAG

Retrieval uses partial index + active file + grep while background indexing is still running.

Context Lens

Gate 1 — pre-send preview of chunks, scores, routing, and governance before LLM inference.

Context pack

Curated markdown knowledge with manifest; activated via @pack:name or packs panel.

Glass box

Transparent routing — provenance chips, routing badge, reasoning timeline, inspector drawer.

Readiness phase

Unified UI state: offline, starting, indexing, ready, limited, error. Gates chat send only for offline/starting/error.

Ollama

Local LLM runtime for embeddings and chat on the default path. Engine manages lifecycle.

BYOK

Bring your own key — optional cloud provider routing with keys in VS Code secrets.

Quality map

Stored governance scan results (clones, structure) surfaced in lens governance drawer.

Assisted action

Structured proposed change with risk level; human confirms before apply.

Image evidence

Screenshot or image file attached to a chat message — paste, drag-drop, or file picker. Normalized, previewed in Context Lens, and sent to a vision-capable model. Default path is local Ollama; not indexed as repo chunks.

Vision route

Which tier will answer an image question: local (Ollama VLM), cloud (explicit BYOK vision model), or needs_vlm (pull a vision model in Engine).

ARGUS

Automated Review for GUI Usability & Standards — structured visual audit on screenshots. Returns severity-tagged findings (including WCAG references). Report only; no auto-fix. Paste path from chat; sandbox path via Playwright in preview worktree.

Visual audit

User-facing action (Run visual audit) that invokes ARGUS on the current image evidence. Distinct from chat Q&A — optimized for findings JSON, not conversational answers.

Support bundle

Export from Engine → About — versions, redacted config, logs for support tickets.

Evidence assembly (v1.0)

Shipped in v1.0 pilot — workspace-agnostic; not tied to one framework layout.

Knowledge objects

Structural facts extracted once at index time (for example RouteObject for API routes) and stored in registry sidecars. The query router answers deterministic structural questions from these objects before running full search.

Retrieve → Expand → Infer

When knowledge lookup misses or the question is semantic, ContextMint runs hybrid search, then expands around hits (file windows, workspace grep, dependency-graph neighbours). Optional LLM evidence selection is last resort only — tier-gated and off by default.

Operational intelligence & reconciliation (v2.0 roadmap)

Planned in v2.0 Engineering Intelligence — see also operational intelligence and Track B. Not shipped in v1.0.

Knowledge reconciliation

When sources disagree (e.g. Git blame, CODEOWNERS, team memory, stale docs), return one scored answer with supporting and conflicting evidence listed — not four similar snippets. Distinct from index reconciliation, which keeps search indexes in sync with the repo.

Operational intelligence retrieval (OIR)

Retrieval scored by relevance and trust, freshness, and evidence coverage — not vector similarity alone. Planned enhancement on the v2.0 Knowledge Fabric; v1.0 ships hybrid search + expand with human gates.

Symbol-aware retrieval

Search returns named artifacts — file, symbol, line range — not anonymous chunks. Ships in v1.0 citations and Context Lens evidence preview.

Track B — enterprise & compliance (roadmap)

Target end of 2027 — extends code intelligence to org-wide knowledge: Jira, Confluence, SharePoint, Git, and partner-selected knowledge bases, with enterprise controls. See Roadmap → Track B.

Track B

Enterprise program: connect Jira, Confluence, SharePoint, Git, and other knowledge bases into one governed index; ACL + multi-tenant isolation; SSO/RBAC admin console; tamper-evident audit; SOC 2 evidence packaging — same retrieval core and human-gate model as v1.0.

Connected knowledge sources

Track B ingests from systems teams already use — tickets (Jira), wikis (Confluence), document libraries (SharePoint), code (Git org-wide), and additional adapters (Notion, internal portals) prioritized with design partners. v1.0 pilots are Git-first; connectors ship when gated.

Knowledge Fabric (v2.0)

Unified graph model for code, memory, ownership, conventions, and decisions — all features query projections over one schema. Ships in v2.0; Track B adds org documents and compliance exports on top.

Verification Engine & roadmap terms

These capabilities are product direction after v1.0 — see Roadmap and Verification Engine. Plain-language overview: Terms explained.

Blast radius

If you change or remove X, what else breaks? ContextMint walks real dependency links (imports, services, features) to list downstream impact — graph traversal, not an LLM guess. Target: under 10 seconds per query at PR review time.

Verification Engine

Future product layer: check every PR or AI change against ownership, blast radius, architecture rules, and policy before merge. v1.0 ships the foundation (Engine, chat, Context Lens, quality map, gates).

Engineering Reality Graph (Fabric)

Unified map of repo + org facts: ownership, dependencies, ADRs/decisions, policies, and drift baselines. Chat and agents cite this graph instead of inventing structure.

Repository DNA

One-page engineering summary — topology, CODEOWNERS, hotspots, conventions — for fast onboarding and review context.

Decision graph

Links architecture decisions (ADRs), PRs, issues, and docs to the code they explain. Answers “why is it built this way?” with verification status per edge.

Drift (architecture drift)

Live code diverging from recorded intent. Bad drift = unintended violation. Accepted drift = deliberate change recorded so alerts stop firing.

MCP (Model Context Protocol)

Open standard for AI clients to call typed tools (e.g. get_blast_radius) with JSON responses. ContextMint's MCP server is planned — read-only graph tools under RBAC; not available in v1.0. See MCP server (roadmap).

Agent governance

Policy and RBAC for AI agents: scoped read tools, queued proposals, audit trail, and mandatory human patch preview before writes. MCP layer on roadmap; Gates 1 & 2 ship in v1.0.

Fabric citation

Machine-readable pointer in chat metadata to a graph node or edge (decision, feature, dependency path) — so answers trace to evidence, not prose alone. Roadmap enhancement to chat SSE.

Sovereign server (deployment profile)

Shared on-prem API + heavy models (e.g. vLLM) on customer GPU; developers use extension only with OIDC. Distinct from laptop local-first (Engine + Ollama). Planned — see engineering blueprint; not Hub71 demo topology.