Skip to content

ROADMAP.md

Latest commit

 

History

History
169 lines (143 loc) · 9.1 KB

ROADMAP.md

File metadata and controls

169 lines (143 loc) · 9.1 KB
Raw

Squadrant — Roadmap

Last updated: 2026-04-06 Status: Active development. Priorities shift based on project deadlines.

Current State

Squadrant v0.1.x is a working multi-project agent orchestration system with:

  • 3-tier hierarchy (Command → Captain → Crew) via cmux + Agent Teams
  • Obsidian hub/spoke vaults for offline status dashboards
  • Session freshness logic (daily + template hash)
  • Self-enhancement via learnings system
  • CLI: init, launch, status, doctor, projects, shutdown, feedback

✅ Shipped

Landed since the entries below were written (as of v0.8.2):

  • Global effort dial — squadrant effort max|balance|low tokenomics dial (#317 / #381).
  • Monorepo reorg — six internal packages (shared/core/agents/workspaces/web/cli) in a one-way DAG, single bundled bin.
  • Daemon-direct delivery — crew/handoff delivery moved onto the daemon→cmux path; notify-relay deleted (#332).
  • Semantic crew heartbeat — CREW IDLE / QUIET / STALLED lifecycle signal (#354).
  • stopped project status + orphan-crew reap — daemon reaps orphaned crews and marks intentional shutdown as stopped (#324 / #323 / #388).
  • Control-plane store hygiene — task-store GC / purge to keep the daemon store bounded.

Roadmap

P0 — Critical (Next 2 weeks) — ✅ ALL DONE

1. Task Master MCP Integration ✅

Why: Rick gives PRDs, Alan needs structured task decomposition. No tool in squadrant breaks down PRDs into dependency-aware tasks today. What:

  • Register task-master-ai as MCP server in squadrant config
  • Captain uses Task Master tools (parse_prd, get_tasks, next_task, expand_task) to decompose PRDs
  • Task Master's tasks.json lives in project root, captain queries it for crew assignments
  • Add to captain-ops skill: "After receiving a PRD or large scope, use Task Master to decompose before spawning crew" Depends on: task-master-ai npm package (installed globally) Status: Installed globally, registered as MCP server, integrated into captain-ops skill. Works via Max subscription — no separate API key needed.

2. squadrant standup Command ✅

Why: Rick expects daily async standups. Currently captain writes daily-log manually and there's no formatted output for sharing. What:

  • New CLI command: squadrant standup [--project <name>] [--all] [--yesterday]
  • Pure bash/TS — zero LLM tokens (inspired by CCPM's pattern)
  • Reads: spoke vault status.md, daily-logs/, git log --since=yesterday per project
  • Output: formatted markdown block (what done, what next, blockers, time allocation)
  • Optional: --slack flag to post to a webhook Effort: ~2-4 hours Status: Implemented in src/commands/standup.ts.

3. Session Handoff Files ✅

Why: New day = forced fresh session = lost context. claude-mem helps but is unstructured. GSD solves this with HANDOFF.json. What:

  • Captain writes handoff.json to spoke vault at session end (manual trigger or auto on shutdown)
  • Schema: { currentState, openBranches, nextSteps, blockedItems, activeTaskMasterTag, decisions }
  • captain-ops startup reads handoff.json if it exists, loads context, then deletes it
  • Command writes hub-level handoff aggregating all projects Effort: ~2-3 hours Status: scripts/write-handoff.sh + scripts/read-handoff.sh. Captain-ops reads on startup, writes on shutdown.

P1 — High Priority (April)

4. Model Routing Config

Why: oh-my-claudecode claims 30-50% token savings by routing exploration→Haiku, planning→Opus, execution→Sonnet. Squadrant currently runs everything on whatever model the session uses. What:

  • Add to config.json: "models": { "exploration": "haiku", "planning": "opus", "execution": "sonnet", "review": "opus" }
  • Captain passes model param when spawning crew via Agent tool
  • Command uses Opus (always) Effort: ~1-2 hours (config + template changes)

5. CI Feedback Reactor Extension — OBSOLETE (reactor engine removed; would need a new auto-delegation mechanism)

Why: Composio AO auto-routes CI failures back to agents. Squadrant reactor detects CI failure but only notifies — doesn't auto-fix. What:

  • New reaction action: auto-fix-ci — on CI failure, re-delegate to captain with failure logs
  • Captain spawns crew specifically for the fix (reads CI output, patches, pushes)
  • Max retry count (default 2) before escalating to command
  • Extend existing reactor architecture, not a new system Effort: ~3-4 hours

6. GSD Integration for Crew

Why: Crew members hit context rot on long tasks. GSD's artifact-driven execution (fresh 200K context per executor, plans carry forward) is the best solution. What:

  • Install GSD as a skill available to crew sessions
  • Crew uses /gsd:do for multi-step implementation tasks
  • Captain passes task specs as GSD-compatible artifacts
  • Crew's subagents get fresh contexts per wave (GSD handles this) Depends on: GSD skill installation

P2 — Medium Priority (May)

7. LLM Wiki / Knowledge Compilation System

Why: Based on Karpathy's viral LLM wiki pattern (Apr 2026). Squadrant's spoke vaults are proto-wikis — captains learn things but don't compile them into cross-referenced, indexed knowledge. See llm-wiki-research-report.pdf in repo root for full analysis. What:

  • Ingest: When captain/crew discovers something notable, it creates/updates wiki pages in spoke vault wiki/ directory
  • Index: wiki/index.md — content-oriented catalog of all wiki pages, updated on each ingest
  • Log: wiki/log.md — append-only chronological record of wiki changes
  • Lint: Periodic health check (orphaned pages, contradictions, stale claims, missing concept pages)
  • Cross-project compilation: Command aggregates spoke wikis into hub-level wiki for cross-project knowledge
  • Query: Any role can search the wiki for relevant context before starting work
  • Hub vault becomes the "compiled knowledge base", spoke vaults are "working wikis"
  • Inspired by: toolboxmd/karpathy-wiki (auto-invocation hooks), rvk7895/llm-knowledge-bases (Obsidian frontend) Architecture mapping (from report):
    • raw/ layer → existing project source code + docs
    • wiki/ layer → new directory in spoke vaults (LLM-maintained)
    • Schema layer → existing CLAUDE.md + squadrant templates
    • Obsidian = IDE, LLM = programmer, wiki = codebase Effort: ~1-2 weeks (phased: basic ingest first, lint later)

8. squadrant retro Command

Why: Rick wants end-of-week/sprint summaries. Currently requires manually reading 5-7 daily logs. What:

  • New CLI command: squadrant retro [--week] [--sprint] [--project <name>] [--all]
  • Aggregates daily logs for the period
  • Groups by: completed, in-progress, blocked, key decisions
  • Computes: tasks completed, PRs merged, crew sessions spawned
  • Output: formatted markdown for sharing Effort: ~2-3 hours

9. Linear Integration — OBSOLETE as specified (depended on the now-removed reactor engine)

Why: Composio AO supports Linear+GitHub+GitLab. Squadrant only has GitHub reactor. Some projects may use Linear for tracking. What:

  • New reactor source: linear-issues (poll via Linear MCP tools already available)
  • Trigger rules for Linear issue state changes, label changes, assignments
  • Action: delegate to captain, same as GitHub flow
  • Bidirectional: captain updates Linear issue status on task completion Effort: ~1 week

P3 — Nice to Have (Future)

10. Web Dashboard

Why: Obsidian requires desktop app. A lightweight web view would allow checking status from phone/browser. What:

  • Simple static page served locally (localhost:3737)
  • Reads same spoke vault status.md files
  • Auto-refresh, similar to Composio AO's dashboard
  • No auth needed (local only)

11. Plugin/Extension System

Why: Composio AO has 7 plugin slots with 21 plugins. Squadrant's architecture is currently fixed. What:

  • Define plugin interfaces: runtime, agent, workspace, tracker, notifier
  • Allow community to add new trackers (Jira, Asana), notifiers (Slack, Discord), runtimes
  • Follow Composio AO's PluginModule<T> pattern

12. Remote VM Support

Why: Originally planned. Rick's Rex system runs on GCP VMs. What:

  • host: "remote" in project config
  • SSH tunnel to remote machine
  • cmux workspace maps to remote tmux session

13. Cost Dashboard

Why: Token costs across multiple projects add up. No visibility today. What:

  • Track token usage per role per project per day
  • Display in squadrant status or web dashboard
  • Alert when daily spend exceeds threshold

Competitor Reference

Tool Stars Watch for
Composio Agent Orchestrator 6K Linear/GitLab tracker plugins, CI feedback loop patterns
GSD (Get Shit Done) 48K Context engineering methodology, wave-based execution
gstack (Garry Tan) 65K Role-based commands, browser integration
oh-my-claudecode 25K Model routing, 19-agent specialization
CCPM 8K Zero-token bash standups, GitHub Issues as source of truth
Claude Task Master 26K PRD decomposition, MCP tools, dependency graphs
1code (YC W26) 5K Trigger agents from GitHub/Linear/Slack mentions