Skip to content

ROADMAP.md

Latest commit

 

History

History
60 lines (43 loc) · 3.62 KB

ROADMAP.md

File metadata and controls

60 lines (43 loc) · 3.62 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: in progress.

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.
  • Let wrappers import candidate tasks, findings, decisions, preferences, incidents, links, and commands without knowing Brigade internals.
  • 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  • Summarize private chat evidence, do not quote raw third-party messages into public docs or handoffs.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: 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.