Skip to content

Latest commit

 

History

History
310 lines (245 loc) · 15.9 KB

File metadata and controls

310 lines (245 loc) · 15.9 KB

LUMI documentation

Documentation for the LUMI VS Code extension and agent workspace (src/, webview-ui/). BroccoliDB has its own docs — do not duplicate substrate content here; link across via AGENT_STACK.md.

Mintlify home · Companion brief · Agent stack · Maintainer · Repository README


Table of contents


At a glance

Workspace-verified figures from papers/companion-brief.md (v1.0.3):

Metric Value Source
Extension ID CardSorting.lumi package.json
Typed tools 62 src/shared/tools.ts
Wired LLM providers 5 src/shared/providers/providers.json
Built-in slash commands 10 src/core/slash-commands/index.ts
Lifecycle hook kinds 8 VALID_HOOK_TYPES in src/core/hooks/utils.ts
Roadmap VS Code settings 5 lumi.roadmap.* in package.json
Read-only tools 12 READ_ONLY_TOOLS in src/shared/tools.ts
Agent modes 2 plan · act

How LUMI differs

LUMI forked the Cline VS Code extension and added governance layers (subagents, BroccoliDB, roadmap gates). For upstream credit, migration notes, and links to Cline docs, see Origins & acknowledgments in the repository README. Full timeline: EVOLUTION.md.

Typical autonomous agent LUMI
Runs until stopped Approval gate per mutating tool
Opaque file changes Diff view before write
Hard to undo Checkpoints after each tool use
“Done” when model says so Completion pipeline + roadmap gates
flowchart TB
  subgraph docs ["docs/ — LUMI session layer"]
    GS[Getting started]
    UG[User guide]
    ARCH[Architecture]
    PAP[Papers]
    API[api/ touchpoints]
  end
  subgraph broc ["broccolidb/docs/ — substrate"]
    BAPI[Public API]
    BCLI[CLI]
    BPAP[Papers]
  end
  GS --> UG
  UG --> ARCH
  ARCH --> PAP
  ARCH --> API
  API -.->|link only| BAPI
  PAP -.->|separate narrative| BPAP
Loading

Product evolution

Cline → DietCode fork → BroccoliDB substrate → governed swarms → LUMI (current). Phase-by-phase history, dated milestones, migration playbook, and legacy inventory:

EVOLUTION.md · README origins


Project configuration

File / directory Doc
.dietcoderules/ Project rules
.dietcoderules/hooks/ Hooks
.dietcodeignore Ignore file
.dietcodeworkflows/ Workflows
ROADMAP.md Roadmap steering · Auto-governance post-mortem

@ mentions: working-with-files


Start here

Doc Description
Home Mintlify landing page
Quick start Install, provider, first task
What is LUMI? Product overview
Agent stack LUMI + BroccoliDB two-layer map
Project map 1-to-1 map of src/ directories
Code ↔ docs Source path → doc page lookup
Maintainer guide CI checks, branding rules, update checklist

Reading paths by audience

Audience Recommended path Time
New user Quick startYour first projectTask management ~30 min
Power user Auto-approve + CheckpointsMCP overviewHooks ~45 min
Team lead Companion briefEvolutionSecurity ~25 min
Designer / PM PhilosophyUser interface design ~25 min
Cline / DietCode migrator EVOLUTION.mdREADME originsProject rules ~20 min
Engineer (agent) WhitepaperArchitecture (current)Project map ~60 min
Privacy / security review SECURITY_BEST_PRACTICES → root README § Local-first ~15 min
Engineer (substrate) BroccoliDB docsRuntime API index varies
Doc contributor MAINTAINERDOCS_GUIDEREWRITE_PLAN ~15 min

Where to document what

Change Update here (docs/) Update in BroccoliDB (broccolidb/docs/)
New LUMI tool or slash command tools-reference/, CODE_TO_DOC_MAP
New wired provider provider-config/README, model guide
Webview UX / approval flow feature or architecture doc
Spider agent ergonomics in IDE api/spider-agent-ergonomics.md cross-link only
AgentContext / capability API link from api/README.md public-api.md, getting-started.md
CLI broccolidb spider cli.md

When in doubt: session behavior → docs/; durable substrate → broccolidb/docs/.


Release & policy

Doc Purpose
MAINTAINER.md When to update docs; CI commands; branding rules
DOCS_GUIDE.md Full documentation map and principles
REWRITE_PLAN.md Rewrite progress and maintenance checklist
SECURITY_BEST_PRACTICES.md Approval gates, ignore files, MCP trust
CODEBASE_STANDARDS.md Repo layout and coding conventions
../SECURITY.md Vulnerability reporting

User guide

Topic Doc
Tasks core-workflows/task-management.mdx
Plan & Act modes core-workflows/plan-and-act.mdx
Files & @ mentions core-workflows/working-with-files.mdx
Slash commands core-workflows/using-commands.mdx
Checkpoints core-workflows/checkpoints.mdx
Tools index tools-reference/README.mdx
All tools tools-reference/all-dietcode-tools.mdx
Model selection core-features/model-selection-guide.mdx
Glossary getting-started/glossary.mdx

Features & customization

Feature Doc
Auto-approve & YOLO features/auto-approve.mdx
Focus chain features/focus-chain.mdx
Subagents features/subagents.mdx
Storage & cache features/storage-and-cache-management.md
Roadmap steering features/roadmap-steering.mdx · Post-mortem
Memory bank features/memory-bank.mdx
Multi-root workspace features/multiroot-workspace.mdx
Hooks customization/hooks.mdx
Skills customization/skills.mdx
Workflows customization/workflows.mdx
Project rules customization/dietcode-rules.mdx
Ignore file customization/dietcodeignore.mdx
MCP mcp/mcp-overview.mdx
Providers (active 5) provider-config/README.mdx

Architecture & internals

Doc Description
Agent stack Two-layer map (LUMI + BroccoliDB)
Architecture (current) Extension module structure
System communication gRPC host bridge, webview messaging
gRPC subscription persistence Persistent streams, idle-timeout incident, runtime architecture
Memory & reasoning Context, cognitive memory tools
Working with subagents Background agent delegation
Roadmap projection quick reference Patch tags, invariants, operator legend — start here
Governed subagent execution Lane modes, SQLite lease authority, wait-for liveness, merge gate, receipts
Governed execution runbook Operator incident playbook, violation catalog, retry flow
Governed execution schema Receipt schema v3 field reference
Governed execution decisions ADR-style design decisions
Governed execution authority SQLite authority, safe reconciliation, deadlock prevention, terminal CAS
Governance Maintainer process, branch protection, release policy
Changelog User-facing release notes (Keep a Changelog)
Support Where to get help
GitHub automation Workflows, templates, and maintainer CI map
Releasing Maintainer marketplace publish runbook
Security best practices Approval gates, ignore files
Central execution funnel One auditable authority for admission, permits, handler dispatch, reliability, and terminal tool events
Task lifecycle authority Transactional generation, cancellation, resume, parent/child propagation, and terminal state
Task cancellation Cancellation request fence and settlement
Task resume and recovery Generation-safe restore and explicit resume
Completion authority migration Historical split-authority migration
Completion funnel Semantic completion and durable handoff to lifecycle
ADR-001 Token Buffer Engine Centralized context optimization, 10-stage DSL compression, vision eviction, and APC prompt cache alignment
ADR-002 Webview State Decoupling Decoupled ChatMessagesContext, WeakMap projection & metrics caching, ChatView modularization, and Cerebras native fetch refactoring
Spider forensic engine Policy/audit layer (BroccoliDB)
User interface design Webview UX patterns

Papers (agent layer)

Doc Audience Purpose
Index All Reading order and two-layer context
Companion brief Leads, evaluators Executive summary · workspace-verified metrics
Philosophy Builders, policy Calm agency · approval · completion gates
Whitepaper Engineers Full technical architecture
Token buffer brief Executives, PMs Token Ingestion Buffer executive summary & 5 debunking proofs
Token buffer philosophy Leads, designers Epistemic compaction & single-turn vision duty
Token buffer whitepaper Engineers, reviewers Formal DSL grammar, Theorem 1–3 proofs & component ablation matrix
MEOW brief Leads, onboarding Executive brief for the Model-Efficient Order-aware Workflow
MEOW philosophy Leads, builders Normative reasoning and calm concurrency principles
MEOW whitepaper Engineers Canonical technical architecture for execution lane swarms
MEOW migration Evaluators Measured evidence and naming evolution

Substrate papers: broccolidb/docs/papers/ — separate narrative.


Runtime API (BroccoliDB touchpoints)

Agent-facing substrate API notes (Spider, snapshots, replay, budgets):

Doc Topic
api/README.md Index
api/spider-agent-ergonomics.md Spider toolkit
api/runtime-snapshots.md Snapshots
api/runtime-replay.md Replay
api/execution-budgets.md Execution budgets

Package docs: broccolidb/docs/README.md


Local development

Mintlify preview:

cd docs && npm install && npm run dev

Quality checks (repository root):

npm run docs:check-all              # all doc guardrails + Mintlify links
npm run docs:check-agent-links      # 24 required docs + link resolution
npm run docs:check-agent-branding   # no stale user-facing DietCode in core dirs
npm run docs:check-root-readme      # README.md parity + live metrics
npm run docs:check-docs-readme      # docs/README.md structure guardrails
npm run docs:check-root-readme-links  # root README relative links
npm run docs:check-readme-metrics     # README + companion-brief vs live codebase
npm run docs:check-links              # Mintlify broken-link pass (included in docs:check-all)
npm run docs:tag-legacy-providers   # after adding unwired provider pages

Doc checks in ci:check-all (except Mintlify-only docs:check-links). Run npm run docs:check-all before publishing docs.


History

Architecture milestone notes (archaeology only): history/architecture/


Principles

  1. 1-to-1 with code — Architecture pages mirror real paths under src/ and webview-ui/.
  2. LUMI user-facing — Product name is LUMI; DietCode only for internal types and legacy filenames.
  3. Two-layer boundary — Session UX docs live here; substrate docs live in broccolidb/docs/.
  4. Measured claims — Metrics cite verifiable paths (package.json, src/shared/tools.ts, etc.).

Calm agency: approve before mutate, checkpoint after tool use, complete only through gates.