Skip to content

Latest commit

 

History

History
163 lines (131 loc) · 9.57 KB

File metadata and controls

163 lines (131 loc) · 9.57 KB

BrainRouter Roadmap

Top-level planning index — kept short. Shipped detail lives in CHANGELOG.md + brainrouter-changelog/; design specs in brainrouter-docs/specs/.


Shipped

Latest published: 0.4.15 (2026-06-22) — the requirement-first workflow + the unified workspace (Chat · Track · Code) with a code-aware Jira-class board, a desktop UI pass, and a closing UX & reliability program (child auto-resume, destructive-command guard, MCP auto-reconnect + brain/tools split, provider gallery with live /models, cross-session usage heatmap, cross-workspace session menu).

Version Theme Date
0.4.15 Requirement-first workflow · unified workspace (Chat·Track·Code) · desktop UX & reliability pass 2026-06-22
0.4.14 Memory accuracy (2 rounds) · grid TUI + fleet sidebar · workers that report back 2026-06-10
0.4.13 Sub-agent result delivery (resume + synthesis guard) · REPL polish 2026-06-07
0.4.12 The Build Loop · multi-agent reconnect + parent-wait timeouts · /queue · accuracy fixes 2026-06-05
0.4.11 Worktree merge-back isolation · self-hydrating config · memory verify + churn decay 2026-06-04
0.4.10 Memory-home hardening · mobile dashboard · Cloudflare Worker runtime 2026-06-03
0.4.9 Dashboard redesign ("The Memory Instrument") · auth refresh tokens · API hardening 2026-06-03
0.4.8 Deterministic multi-phase workflow orchestration (PhasePlan engine) 2026-06-02
0.4.7 Coding-agent parity · agentic autonomy · staleness-aware recall 2026-06-02
0.4.6 Behavior-preserving structural refactor 2026-06-01
0.4.5 Release hygiene · competitive catch-up · CLI polish 2026-06-01
≤ 0.4.4 Memory depth + pipeline + multi-agent foundations → changelog

Full detail: brainrouter-changelog/.


0.4.14 — Memory Accuracy + TUI shell · branch release/0.4.14 · spec memory-accuracy.md

Benchmark-driven recall overhaul (MemBench · LoCoMo · LongMemEval), in two rounds. Status: SHIPPED 2026-06-10 — merged to main, tagged v0.4.14, published to npm.

Round 1 — granularity (shipped): the one-record-per-session granularity wrecked long-session reranking/judging. Fixed by chunking on import, a length-aware reranker cap, a judge result-floor, embed-on-import, and a transient-embed retry.

  • MEM-AUDIT · MEM-CHUNK · MEM-RERANK · MEM-JUDGE · MEM-VEC · MEM-EMBED-RETRY · ASYNC-1

Round 2 — recall quality (shipped): the clean 6-split sweep exposed that the reranker/judge replaced the retriever order — collapsing recall and losing to a plain recency baseline. Fixes follow one rule: score → sort → take top-N, never hard-drop. Results: reports/0.4.14-recall-delta.md.

  • MEM-JUDGE2 — judge reorders (approved-first), never drops below the retriever
  • MEM-BLEND — blend reranker score with the recency/RRF score instead of replacing it
  • MEM-RERANK2 — char-budgeted reranker pool to cut long-session latency
  • MEM-ROUTE — reflective queries skip the cross-encoder (retriever+judge path)
  • MEM-EVAL — per-stage benchmark gate + final 6-split sweep vs the saved baseline

Deferred per ADR-001: ASYNC-2/3 (worker-thread store). Follow-up: benchmark importer per-record timestamps so the recency signal is measurable.

TUI shell + worker orchestration (merged to the release branch 2026-06-09, #344): grid TUI (view router · 70/30 workspace split · context sidebar with the running fleet split into sub-agents / workers / workflows — role, worktree marker, live elapsed), model-spawned background worker threads (depth/tier-gated), a completion inbox so detached background work reports its results into the agent's next turn, and a scrollback overhaul (accurate height packing, scroll mode, live turn timer).


0.4.15 — The Agent Behind the Glass · plan brainrouter-roadmap/0.4.15.md

A complete reverse-engineering of the strongest reference coding agent — decomposed into twelve subsystems, each mapped against what BrainRouter has and rebuilt as testable contracts. Keeps the 0.4.14 UI. Threads: A chat ergonomics · B prompt OS + turn-engine contracts (deliverable guard, batching, autonomy/ assessment rules, denied-tool semantics, verification gate) · C per-tool contracts (read-before-edit, unique-match edits) · D context OS (stable cached prefix, reminder channel, structured compaction) · E first-class task tracker · F programmable workflows (scripted stages, schema-validated child outputs, budgets, resume) · G background runtime (bg shell + output polling, wait-until, cron agents, recurring loop) · H session lifecycle (resume, rewind, branch, chapters) · I plan mode + declarative permissions + hook gates · J interaction layer (structured multi-choice questions, side-task chips, skill auto-trigger, md commands) · K headless stream-json surface · L the agent-behavior benchmark that gates it all. · M BrainRouter Desktop — native macOS/Windows app (Electron + React over the unchanged agent runtime; chat, approvals, fleet sidebar, sessions as real UI; signed installers + auto-update; ships as alpha beside the TUI).


Delivered in 0.4.15 — The unified workspace: Chat · Track · Code

One app, three modes over the same workspace + the same memory, so planning, coordination, and implementation never leave the building. A mode switcher (top of the left sidebar) flips the surface; the project, the memory, and the agent are shared across all three. Shipped in 0.4.15.

  • Chat — conversational/assistant mode: ask, explore, design, and draft without the full coding-agent loop. Lightweight and read-only (the agent is pinned to look-only access — no writes or shell); promotes a thread into a Code turn or a Track item when it's ready.
  • Track — a first-party, Jira-class project-management surface, fully in-app so everything stays in one place. Each workspace is a project with its own management. Target scope (everything an engineering team needs):
    • Work items — issues / stories / bugs / tasks / sub-tasks, epics, and custom types; configurable workflow states + transitions; priority, labels, components, assignees, watchers, estimates, due dates, comments, attachments, activity history.
    • Planningbacklog, sprints/iterations (start/complete, capacity, burndown), boards (kanban + scrum, swimlanes, WIP limits), and roadmap (epics over time, dependencies).
    • Views & query — board / list / table / timeline, saved filters, and a JQL-style query language; reports (velocity, cumulative flow, cycle/lead time).
    • Automation & process — rules/triggers, templates, required fields/gates, SLAs; per-project configuration and permissions.
    • Code-aware (the differentiator) — the agent reads and writes the tracker as a first-class tool: items link to branches / commits / PRs / review findings / artifacts; status moves with the work; the existing requirement → plan → task → review → verify flow maps onto Track items; everything captures into BrainRouter memory with full provenance. Each project's board is queryable by the agent so it can pick up, update, and close work itself.
  • Code — today's agentic coding mode (the desktop + TUI), unchanged.

Shared substrate: one memory pipeline, one workspace/session model, one provider config; Track items, Chat threads, and Code turns all reference the same provenance-tracked records. Delivered across the full stack: data model + per-workspace store (packages/types/track.ts + packages/core/src/track/), the track_query/track_update agent tools, the /track CLI, and the desktop Track surface (nine views — Board · List · Backlog · Sprint · Roadmap · Reports · Automation · Members · Sync) with a JQL-style query language, automation rules, per-project roles/permissions, and two-way GitHub Issues sync.


0.5.0 — Power-user surface

  • Fullscreen /focus TUI + plugin marketplace + trust/signature model.
  • Brain Phase 6 — engineering sync providers (Git / GitHub / docs / logs) + proactive situation reports.
  • Dashboard per-agent recall view + drill-down.

Designed, not scheduled

Wishlist (post-0.5.0)

  • Docker image for the MCP server.
  • Dashboard memory explorer with recall-score explanations.
  • Dashboard parity with CLI goals / hooks / workflows.
  • Verified provider matrix (OpenAI, Anthropic, Gemini, OpenRouter, LM Studio, Ollama).
  • @kinqs/brainrouter-sdk 1.0 public API lock.

Documentation map

File Purpose
CHANGELOG.md Shipped changes
brainrouter-changelog/ Per-version changelog detail
brainrouter-roadmap/ Per-release plans
brainrouter-docs/specs/ Design specs (memory-accuracy, build-loop, per-session isolation, multi-agent result delivery)
brainrouter-docs/decisions/ ADRs (async store worker)
brainrouter-benchmark/reports/ Published benchmark results