feat(research): comparative study Pi Agent + pi-subagents + devstack vs RooSync/Claude Code harness

[jsboige]

Summary

Comparative study of 3 alternative/open-source agent harness ecosystems against our current stack (RooSync v2.3 + Claude Code + Roo Code + roo-state-manager MCP + 6-machine fleet). Goal: identify migration opportunities, adaptation patterns, and architectural inspiration.

Projects Under Study

1. Pi Agent — 56.8k stars, MIT, v0.76.0

TypeScript mono-repo coding agent toolkit. Key components:

• pi-ai: Unified multi-provider LLM API (Claude, GPT, Gemini, Bedrock, Vertex, OpenRouter, local models)
• pi-agent-core: Agent runtime, tool calling, state management
• pi-coding-agent: CLI coding harness (TUI)
• pi-tui: Terminal UI

Notable features:

• Non-blocking UI (model switching, effort, status while running)
• Advanced session history (/tree, /fork, /clone for time-travel/branching)
• Hot-reload extensions (/reload without restart)
• Designed for extreme customizability ("tell pi to change itself")

2. pi-subagents — 1.6k stars

Extension for Pi with async subagent delegation. Strong parallels to our architecture:

• 8 built-in agents: scout, researcher, planner, worker, reviewer, oracle, delegate, custom
• Execution modes: parallel, chain, background (like our executor/meta patterns)
• Worktree isolation per subagent (we do this via .claude/worktrees/)
• Intercom bridge for cross-agent communication (like our dashboard workspace)
• Recursion guards and skill injection
• Delegation patterns resemble our Roo→Claude coordination

3. devstack — 21 stars

"Agentic practices knowledge base + supporting software toolkit" — opinionated guide for agentic programming. Key insights:

• Karpathy LLM Wiki pattern: structured knowledge base with inbox/ → sources/ → wiki/ pipeline (we use MEMORY.md + codebase_search)
• Extension ecosystem curation: pi-packages.json manifest for fleet-wide consistent setups
• Context management tools: pi-context-prune (recoverable summarization), pi-vcc (zero-LLM algorithmic compaction), pi-continue-after-compaction (auto-continue post-compaction)
• pi-multiloop: Battle-tested autoloop for long-running agent sessions (analogous to our executor cycles)
• pi-multicodex: Automatic account rotation for quota/rate limits
• Model evaluations: Comparative rankings of frontier + open models for coding tasks

Comparison Axes

A. Architecture

Aspect	Our Stack	Pi + Extensions
Agent runtime	Claude Code (Anthropic) + Roo Code (extension)	pi-agent-core (customizable TypeScript)
LLM providers	Claude API + GLM/z.ai + vLLM	Unified pi-ai (Claude, GPT, Gemini, Vertex, Bedrock, OpenRouter, local)
Multi-machine	RooSync GDrive + 34-tool MCP	Single-machine focus (subagents local)
Coordination	Dashboard workspace + RooSync messages	Intercom bridge (local subagents)
State persistence	SQLite + Qdrant + GDrive	Session history with fork/clone
Context management	Auto-compaction (90% threshold)	pi-vcc (algorithmic), pi-context-prune (recoverable)

B. What They Do Better

1. LLM provider abstraction: pi-ai unified API vs our ad-hoc per-provider config
2. Non-blocking UI: Pi allows model/status operations while running
3. Session branching: /fork, /clone, /tree — we have no equivalent
4. Zero-LLM compaction: pi-vcc uses deterministic extraction, no summarizer model needed (avoids 400 errors on long sessions)
5. Recoverable pruning: pi-context-prune keeps originals retrievable on demand
6. Extension hot-reload: /reload without restart vs our VS Code restart requirement
7. Account rotation: pi-multicodex handles multi-account quota automatically
8. Auto-continue post-compaction: pi-continue-after-compaction watches and sends continue

C. What We Do Better

1. Multi-machine fleet coordination: 6 machines with RooSync GDrive sync, heartbeat, dashboard workspace — Pi is single-machine
2. Cross-machine task dispatching: Hash-based partitioning (ROO_FLEET_ROSTER), claim discipline, anti-double-claim
3. MCP ecosystem: 34-tool roo-state-manager for coordination, semantic search, conversation indexing
4. Scheduler integration: schtasks + worker scripts for autonomous cycles
5. Agent claim discipline: Verified claims, SHA verification, pre-claim anti-overlap
6. CI/CD pipeline: check-submodule-pointer, validation scripts, PR mandatory rules

D. Migration/Adaptation Opportunities

Opportunity	Source	Effort	Impact
Unified LLM provider layer	pi-ai pattern	Medium	Simplify multi-provider config across fleet
Zero-LLM compaction	pi-vcc pattern	Medium	Eliminate summarizer 400 errors, deterministic
Recoverable context pruning	pi-context-prune	Medium	Lossless compaction with on-demand retrieval
Session branching/forking	Pi /fork//clone	High	Enable experimentation without context loss
Extension hot-reload	Pi /reload	Low (if Pi adopted)	Eliminate VS Code restart cycle
Auto-continue post-compaction	pi-continue-after-compaction	Low	Prevent stalls after auto-compaction
Account rotation	pi-multicodex pattern	Low	Multi-key rotation for API quotas
Worktree per-subagent	pi-subagents	Medium	Stronger isolation for parallel agents
Fleet extension manifest	devstack pi-packages.json	Low	Consistent MCP/extension setup across 6 machines
Karpathy wiki pattern	devstack inbox→sources→wiki	Low	Structured knowledge management

Proposed Approach

1. Deep-dive pi-subagents delegation patterns — closest to our multi-agent architecture, evaluate for Roo Code integration
2. Prototype zero-LLM compaction — pi-vcc algorithmic approach for roo-state-manager
3. Evaluate pi-multiloop — compare with our executor cycle pattern, identify improvements
4. Study fleet-wide manifest — pi-packages.json concept for 6-machine config consistency
5. Assess Pi as alternative harness — for executor machines currently using Claude Code

Context

• Current stack: RooSync v2.3 + Claude Code + Roo Code + roo-state-manager (34 MCP tools) + 6-machine fleet
• Our executor cycles: 360min intervals, auto-stop after 3 IDLE, wake routing via [WAKE-CLAUDE]
• Compaction: 200k/90% universal threshold, auto-condensation at 92% dashboard

Acceptance Criteria

• Comparative analysis document created in docs/harness/reference/
• Top 3 adaptation opportunities identified with concrete implementation plans
• Decision documented: adopt / adapt / defer for each opportunity
• If adopting: issues created for implementation tasks

/label enhancement

[github-actions]

Missing labels detected:

• type: bug, enhancement, or documentation
• agent: claude-only or roo-schedulable

Please add the missing labels so Project #67 fields are set correctly.

---

Automated by sync-project.yml (#1835)

[myia-ai-01]

[CLAIMED] by claude on myia-ai-01 at 2026-05-28T17:08:29.1636617+02:00