Skip to content

ROADMAP.md

Latest commit

 

History

History
98 lines (73 loc) · 7.6 KB

ROADMAP.md

File metadata and controls

98 lines (73 loc) · 7.6 KB
Raw

Brigade Roadmap

Brigade is being built as a practical daily workflow first, then a portable setup other people can adapt. The core direction is an organized version of real agent work: one command to start, predictable local artifacts, reviewable memory handoffs, and enough inspection to trust the loop during normal work.

Current Phase: Daily Driver

Status: in progress.

  • Local dogfood defaults live in gitignored .brigade/dogfood.toml.
  • brigade work bootstrap prepares a repo for the daily loop.
  • brigade work brief is the start-of-day entrypoint.
  • brigade work run wraps a dogfood run in local work-session artifacts and handoffs.
  • brigade work tasks plus brigade work task add/show/done provide a gitignored local task ledger.
  • brigade work run --queue-next queues extracted follow-up work without duplicating equivalent pending tasks.
  • brigade work import add/list/show/promote gives scanners and wrappers a stable local inbox for candidate work.

Next Phase: Scanner-Ready Inbox

Status: active.

Goal: make Brigade a safe target for local automations that discover useful work.

  • Keep raw scanner output private and gitignored under .brigade/work/imports/.
  • Normalize imports into small records with kind, source, text, timestamps, and metadata.
  • Document the scanner JSONL contract so external producers can target Brigade without importing Brigade internals.
  • Validate and ingest scanner-authored JSONL files.
  • Let wrappers import candidate tasks, findings, decisions, preferences, incidents, links, and commands without knowing Brigade internals.
  • Convert memory-care refresh queues into local task imports.
  • Promote selected imports into the work task ledger, with source metadata preserved.
  • Dismiss noisy imports so scanners can be useful without leaving permanent queue clutter.
  • Batch-promote reviewed imports by source and kind.
  • Surface pending imports and grouped counts in brigade work brief so discovered work appears in the daily flow.

Later Phase: Chat Surface Scanners

Goal: support the common places agent work happens without making any one chat product mandatory.

  • Build adapters for Discord, Slack, Telegram, and export-based chat archives as separate scanner layers.
  • Convert surface-specific events into the local import inbox instead of writing memory directly.
  • Summarize private chat evidence, do not quote raw third-party messages into public docs or handoffs.
  • Use promotion gates so only reviewed, durable, or actionable items become tasks or memory handoffs.
  • Keep source metadata such as workspace, channel, thread, message range, and confidence local unless explicitly exported.

Later Phase: Cybersecurity Plugin

Goal: ship a Brigade cybersecurity plugin with broad coverage for agent workspaces, then go deeper on Brigade's multi-harness, memory, scanner, and dogfood workflows.

Baseline coverage targets:

  • Scan agent workspace configs for hardcoded secrets, exposed tokens, private keys, database URLs, and unsafe environment-variable handling.
  • Audit tool permissions for broad mutable access, wildcard shell access, missing deny lists, dangerous flags, destructive git commands, and unrestricted network commands.
  • Analyze hooks and startup automation for command injection, remote execution, data exfiltration, silent failures, package installs, container escape, reverse shells, clipboard access, log tampering, and persistence behaviors.
  • Audit MCP server configs for high-risk server types, remote transports, shell metacharacters, unpinned npx usage, hardcoded env secrets, sensitive file args, excessive server counts, missing timeouts, and auto-approve behavior.
  • Review agent prompts, skills, subagents, slash commands, and workspace instructions for prompt-injection patterns, hidden instructions, URL execution, data harvesting, output suppression, time bombs, and unsafe auto-run language.
  • Emit graded reports with severity, category scores, evidence snippets, suggested fixes, JSON output, markdown output, HTML or bundle output, and CI-friendly exit codes. Status: started with redacted JSON and Markdown evidence bundles.
  • Support CLI use, GitHub Action use, and local evidence packs.

Brigade-specific additions:

  • Scan Claude Code, Codex, OpenCode, Gemini, Hermes, OpenClaw, VS Code, Zed, dmux, and generic repo-local agent harness surfaces with explicit runtime-confidence labels.
  • Understand Brigade installs: .brigade/, .codex/, .claude/, memory handoff inboxes, roster files, dogfood configs, run artifacts, work imports, memory-care decay files, and public template folders.
  • Treat public-template findings differently from active runtime findings so docs and starter templates do not score like live credentials or enabled tools.
  • Integrate with brigade doctor as a security station and with brigade work import so findings can become reviewable local tasks instead of only console output.
  • Provide safe auto-fix only for narrow cases such as replacing obvious hardcoded sample secrets, tightening generated allow-list examples, or adding missing ignore rules.
  • Produce Memory Handoffs for durable security findings while keeping raw secret evidence redacted.
  • Add policy packs for personal dogfooding, public-repo release checks, CI gates, and strict enterprise workspaces. Status: started with personal, public-repo, and strict.
  • Include dependency and package-manager hardening checks for agent plugin ecosystems, MCP packages, skills, and local tool wrappers.
  • Track false-positive taxonomy, runtime-confidence rules, suppressions, and regression fixtures as first-class project artifacts.

First build slice:

  • Create a plugin scaffold and security scan contract. Status: started with built-in security station, brigade security init, and brigade security scan.
  • Start with config discovery and read-only reporting for Brigade, Claude Code, Codex, and MCP config files. Status: started.
  • Add core rule categories for secrets, permissions, hooks, MCP servers, supply-chain patterns, and agent instructions. Status: started.
  • Output JSON plus readable text, redacted evidence bundles, then route selected findings into brigade work import. Status: started with --output-dir and --import-findings.
  • Keep all raw findings local and gitignored unless the operator explicitly exports an evidence pack. Status: current default.
  • Add local policy defaults, stable finding fingerprints, and suppressions. Status: started with .brigade/security.toml.

Later Phase: Memory Card Decay And Refresh

Goal: prevent durable memory from silently rotting.

  • Track freshness metadata, confidence, evidence, and review dates for memory cards.
  • Run memory-care scanners that detect expired, stale, contradictory, or undersourced cards.
  • Import refresh candidates into Brigade as local work imports.
  • Promote refresh candidates into tasks or memory handoffs after review.
  • Auto-fix only within safe gates where source evidence is current, low-risk, and locally reviewable.
  • Treat bootstrap truncation as a hard failure. Bootstrap files stay slim, cards hold durable detail, and doctor checks enforce the boundary.

Later Phase: Portable Operator Setup

Goal: keep the system usable by the original operator while making it adaptable by others.

  • Keep Codex-first defaults, with Claude Code, OpenCode, Hermes, OpenClaw, and generic harness paths supported through writer-specific inboxes.
  • Make local paths configurable and gitignored.
  • Provide templates for fresh-start users without publishing private workspace state.
  • Keep public repo docs focused on patterns, commands, and safety contracts.
  • Leave release, tag, push-to-main, and production-impacting actions behind explicit approval gates.