Shared Claim Board

This board coordinates Pi agents and external agents.

All agents must check this page and ../../log.md before taking work.

External agents are standalone operators. They should claim and complete their own chunk without managing Pi-specific runtime details unless explicitly assigned to a Pi/runtime setup task.

Active Claims

Coordinator correction: after the pivot to Pi workcell architecture research, Pi agents must hold distinct research questions. External agents should treat the active claims below as authoritative unless released or stale.

RELEASE: Q-00 | lead | 2026-05-11T02:50:00Z

Status: ANSWERED Task: Architecture Synthesis Stage: T0-T2 Complete Output: /workcell/llm-wiki/wiki/research/answers/q-00-architecture-synthesis.md Summary: Recommended minimal Pi workcell architecture: Pi Teams framework, single Docker container, LLM-wiki file-based knowledge depot, T0-T7 workflow. Avoids complex frameworks and Python dependencies.

CLAIM: Q-00 | lead | 2026-05-11T01:52:00Z

Status: CLAIMED Lease: 90 minutes Task: Architecture Synthesis Stage: T0/T1 Scope: Coordinate question assignment and synthesize answered research into the final recommended Pi workcell architecture Expected output: /workcell/llm-wiki/wiki/research/answers/q-00-architecture-synthesis.md

RELEASE: Q-05 | researcher | 2026-05-11T02:00:00Z

Status: ANSWERED Task: Knowledge Depot / RAG Architecture Research Stage: T1 Review/Research Output: /workcell/llm-wiki/wiki/research/answers/q-05-knowledge-depot-rag.md Summary: Recommended LLM-wiki as primary knowledge depot with optional SQLite enhancement, avoiding complex RAG systems for MVP.

Active Claims

RELEASE: Q-04 | researcher | 2026-05-11T02:20:00Z

Status: ANSWERED Task: Work Shape / Lifecycle Research Stage: T1 Review/Research Output: /workcell/llm-wiki/wiki/research/answers/q-04-work-shape-lifecycle.md Summary: Comprehensive T0-T7 lifecycle definition with stage details, mandatory artifacts, decision points, coordination rules, and artifact standards.

CLAIM: Q-06 | builder-reviewer | 2026-05-11T01:47:00Z

Status: HANDLED Lease: 90 minutes Task: Runtime Container Shape Research Stage: T1 (Research) → COMPLETE Scope: Answer Q-06 about Docker vs pi-container-sandbox vs separate containers Expected output: /workcell/llm-wiki/wiki/research/answers/q-06-runtime-container-shape.md

RELEASE: Q-06 | builder-reviewer | 2026-05-11T02:10:00Z

Status: ANSWERED Output: /workcell/llm-wiki/wiki/research/answers/q-06-runtime-container-shape.md Next: Available for architecture synthesis (Q-00)

RELEASE: Q-01 | researcher | 2026-05-11T03:05:00Z

Status: ALREADY ANSWERED Task: Pi Teams Fit Research Output: /workcell/llm-wiki/wiki/research/answers/q-01-pi-teams-fit.md (by opencode/ext-agent) Summary: Comprehensive analysis confirming pi-teams is strong fit for workcell. Identifies features, gaps, recommendations, and risks. No framework replacement needed for MVP.

RELEASE: Q-03 | researcher | 2026-05-11T03:15:00Z

Status: ALREADY ANSWERED Task: Framework Comparison Research Output: /workcell/llm-wiki/wiki/research/answers/q-03-framework-comparison.md (by opencode/ext-agent) Summary: Comprehensive framework comparison confirming pi-teams is best choice for MVP. Rejects CrewAI/AutoGen due to Python complexity. Considers PydanticAI for future typed-output validation.

Claim Format

Append claims here and to ../../log.md.

### CLAIM: <task-id> | <agent/session> | <timestamp UTC>

Status: CLAIMED
Lease: 90 minutes unless refreshed
Task: <issue/chunk/path>
Stage: T0/T1/T2/T3/T4/T5/T6/T7
Scope: <what this session will answer/do>
Expected output: <wiki path / Forgejo comment / patch>

Check-In Format

### CHECK-IN: <task-id> | <agent/session> | <timestamp UTC>

Status: CLAIMED
Progress: <brief>
Next: <brief>

Release Format

### RELEASE: <task-id> | <agent/session> | <timestamp UTC>

Status: HANDLED / PARTIAL / BLOCKED / RELEASED
Output: <links>
Next: <follow-up>

Stale Claim Rule

A claim is stale after 90 minutes without check-in. A later agent may take it over only if it cites the stale claim and writes a new claim.

CLAIM: Q-01 | opencode/ext-agent | 2026-05-11T01:49:12Z

Status: CLAIMED Lease: 90 minutes unless refreshed Task: Q-01 Pi Teams Fit — evaluate features, gaps, role config, and task flow Stage: T1 Review / Research Scope: Research pi-teams capabilities from repo/docs/container runtime, evaluate fit for d3-tui-triad workcell (lead/researcher/builder-reviewer), document strengths and gaps Expected output: llm-wiki/wiki/research/answers/q-01-pi-teams-fit.md

RELEASE: Q-01 | opencode/ext-agent | 2026-05-11T02:10:00Z

Status: HANDLED Output: llm-wiki/wiki/research/answers/q-01-pi-teams-fit.md Summary: Evaluated pi-teams features, gaps, and fit for d3-tui-triad. Core team/task/messaging/autonomous-polling surface is a strong fit. Gaps (task lifecycle enforcement, external agents, Forgejo) are addressable via prompt discipline + LLM-wiki. No alternative framework needed. Next: Mehdi's decisions on plan approval mode, quality gate hooks, agent models, and predefined team extraction.

CLAIM: Q-02 | opencode/ext-agent | 2026-05-11T01:54:47Z

Status: CLAIMED Lease: 90 minutes Task: Q-02 LangGraph Fit — evaluate whether LangGraph improves d3-tui-triad beyond what pi-teams provides Stage: T1 Review / Research Scope: Research LangGraph state machines, durable execution, human-in-the-loop, Pi resource impact, and T0-T7 lifecycle mapping Expected output: llm-wiki/wiki/research/answers/q-02-langgraph-fit.md

RELEASE: Q-02 | opencode/ext-agent | 2026-05-11T02:15:00Z

Status: HANDLED Output: llm-wiki/wiki/research/answers/q-02-langgraph-fit.md Summary: LangGraph not recommended for MVP. Would add Python runtime, orchestration collision with pi-teams, and 2-3 sessions of pre-staging. pi-teams plan approval mode provides the T0-T3 gate with zero new infrastructure. Defer LangGraph; revisit only if prompt-discipline coordination consistently fails. Next: Enable plan approval mode as the direct alternative.

CLAIM: Q-11 | codex/ext-agent | 2026-05-11T02:47:00Z

Status: CLAIMED Lease: 90 minutes unless refreshed Task: Q-11 Remote UI / Observability — evaluate pi-gui, pi-agent-dashboard, Forgejo UI, tmux, and health-check surface Stage: T1 Review / Research Scope: Research all remote viewing/control options, test what works on current relik-pi4 setup, recommend simplest stack Expected output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-11-remote-ui-observability.md

CLAIM: Q-03 | opencode/ext-agent | 2026-05-11T02:55:00Z

Status: CLAIMED Lease: 90 minutes Task: Q-03 Framework Comparison — eval CrewAI, AutoGen, PydanticAI vs pi-teams Stage: T1 Review / Research Scope: Compare frameworks for improvisation+order, setup friction, Pi/tmux/Forgejo/LLM-wiki integration, MVP rejection criteria Expected output: llm-wiki/wiki/research/answers/q-03-framework-comparison.md

CLAIM: Q-09 | opencode/ext-agent | 2026-05-11T00:05:00Z

Status: CLAIMED Lease: 90 minutes unless refreshed Task: Q-09 Bun / Pi Install / Model Routing — evaluate install/runtime pattern for Pi agents on RPi4 hardware Stage: T1 Review / Research Scope: Research Bun vs Node for Pi CLI, npm install patterns, Kimi/MiniMax key routing, image-baked vs mounted config, current container runtime analysis Expected output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-09-bun-pi-install-model-routing.md

RELEASE: Q-11 | codex/ext-agent | 2026-05-11T02:50:00Z

Status: ANSWERED Task: Remote UI / Observability — evaluate pi-gui, pi-agent-dashboard, Forgejo UI, tmux, and health-check surface Stage: T1 Complete Output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-11-remote-ui-observability.md Summary: SSH + tmux is the primary surface today. Forgejo web UI is the secondary repo ledger (requires container). pi-agent-dashboard is aspirational (requires server + bridge wiring). pi-gui is the wrong tool — local-only pi CLI wrapper, not remote monitoring. All Docker containers crashed under load; restart only when safe. Next: Run docker start from-forgejo when system load permits; wire pi-agent-dashboard bridge extension into agent pi settings if browser UI is desired.

CLAIM: Q-03 | opencode/ext-agent | 2026-05-11T03:30:00Z

Status: CLAIMED Lease: 90 minutes unless refreshed Task: Q-03 Framework Comparison — compare CrewAI, AutoGen/Microsoft Agent Framework, PydanticAI against pi-teams for d3-tui-triad workcell Stage: T1 Review / Research Scope: Evaluate each framework on improvisation+order, setup friction, integration with pi-teams/tmux/Forgejo/LLM-wiki, and Pi resource fit. Identify which, if any, should be used for typed outputs/planning/evaluation only rather than full orchestration. Expected output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-03-framework-comparison.md

CLAIM: Q-03 | opencode/ext-agent | 2026-05-11T01:58:38Z

Status: CLAIMED Lease: 90 minutes Task: Q-03 Framework Comparison — evaluate CrewAI, AutoGen/Microsoft Agent Framework, Pydantic AI vs pi-teams

RELEASE: Q-03 | opencode/ext-agent | 2026-05-11T02:20:00Z

Status: HANDLED Output: llm-wiki/wiki/research/answers/q-03-framework-comparison.md Summary: CrewAI, MS Agent Framework, and Pydantic AI all rejected for MVP. All three are Python-based, lack terminal GUI, and are designed for service automation — not live coding sessions. pi-teams remains the best fit. Enable plan approval mode as the alternative to framework replacement. Next: Test plan approval mode to validate it covers the T0-T3 gate.

CLAIM: Q-12 | opencode/ext-agent | 2026-05-11T03:05:00Z

Status: CLAIMED Lease: 90 minutes unless refreshed Task: Q-12 External Agent Coordination — define prompt, protocol, and coordination rules for external agents Stage: T1 Review / Research Scope: Define exact external agent prompt, onboarding sequence, claim/collision avoidance, Pi-specific boundaries, answer format standards, and completion criteria Expected output: llm-wiki/wiki/research/answers/q-12-external-agent-coordination.md

RELEASE: Q-03 | opencode/ext-agent | 2026-05-11T03:10:00Z

Status: ANSWERED Output: llm-wiki/wiki/research/answers/q-03-framework-comparison.md Summary: Reject CrewAI and AutoGen for MVP (Python-heavy, would discard working pi-teams/tmux setup). Keep pi-teams as foundation. Consider PydanticAI narrowly for typed-output validation at T5-T6 stage. Quick win: extract pi-teams prompts to .pi/agents/*.md + teams.yaml (~15 min). Next: Available for other OPEN questions.

CLAIM: Q-04 | opencode/ext-agent | 2026-05-11T02:22:00Z

Status: CLAIMED Task: Q-04 Work Shape / Lifecycle — define T0-T7 artifacts, gates, and coordination

RELEASE: Q-04 | opencode/ext-agent | 2026-05-11T02:30:00Z

Status: HANDLED Output: llm-wiki/wiki/research/answers/q-04-work-shape-lifecycle.md Summary: Full T0-T7 lifecycle defined with mandatory artifacts per stage, multi-agent agreement protocol, external agent integration rules, duration caps, and escalation triggers. Builds on existing task-lifecycle.md and agent-role-contract.md. Recommends plan approval mode as mechanical T3→T4 gate.

RELEASE: Q-09 | opencode/ext-agent | 2026-05-11T00:15:00Z

Status: ANSWERED Output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-09-bun-pi-install-model-routing.md Summary: Keep Node.js for Pi CLI. @mariozechner/pi-coding-agent@0.73.1 is deprecated → migrate to @earendil-works/pi-coding-agent. Model keys should stay mounted, not baked. Bun stays host-only. Next: Mehdi decisions on (1) package migration to @earendil-works, (2) per-agent vs shared API keys, (3) Pi CLI on host for testing.

CLAIM: Q-13 | codex/ext-agent | 2026-05-11T02:55:00Z

Status: CLAIMED Lease: 90 minutes unless refreshed Task: Q-13 Viable Remote Monitoring GUI — research ttyd, Tailscale Serve, static status pages, and other lightweight options Stage: T1 Review / Research Scope: Evaluate every browser-accessible monitoring option that works on Pi 4 without adding Docker containers, test what is installable, recommend the simplest viable stack Expected output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-13-viable-remote-gui.md

CLAIM: Q-07 | opencode/ext-agent | 2026-05-11T03:12:00Z

Status: CLAIMED Lease: 90 minutes Task: Q-07 Toolchain / KOS Contract — best way to expose KOS/Shachi/ORCA capabilities to agents Stage: T1 Review / Research Scope: Research d3-tui's KOS dependency, toolchain contract, what broke previously, and what docs/smoke tests are needed Expected output: llm-wiki/wiki/research/answers/q-07-toolchain-kos-contract.md

RELEASE: Q-13 | codex/ext-agent | 2026-05-11T03:00:00Z

Status: ANSWERED Task: Viable Remote Monitoring GUI — research lightweight browser-accessible options Stage: T1 Complete Output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-13-viable-remote-gui.md Summary: Three-tier recommendation. Tier 1 (now): Tailscale Serve + Python HTTP server + static HTML status page — zero install, ~20MB RAM, shows agent state, claims, wiki. Tier 2 (later): Forgejo native behind Tailscale Serve for full repo/issues browser. Tier 3 (optional): ttyd + tmux for live agent panes in browser. All over tailnet HTTPS — no Docker needed for GUI. Next: Deploy Tier 1 — write status generator, serve via Tailscale, verify from Mac browser.

RELEASE: Q-04 | opencode/ext-agent | 2026-05-11T02:40:00Z

Status: HANDLED Output: llm-wiki/wiki/research/answers/q-04-work-shape-lifecycle.md Summary: Revised lifecycle based on wild research. Only 2 hard gates (plan approval, validation). Rest is lead checklist. Claude Code teams, SWE-agent, and OpenHands all use simpler patterns — the d3-tui-triad is already more structured than any production system.

RELEASE: Q-03 | opencode/ext-agent | 2026-05-11T04:00:00Z

Status: ANSWERED Output: /mnt/kitchen/from-house/workspace/d3-tui-pi-teams-proto/llm-wiki/wiki/research/answers/q-03-framework-comparison.md Summary: CrewAI and AutoGen rejected for MVP (Python, replace pi-teams, lose tmux, too heavy for Pi). PydanticAI deferred as complementary T5/T6 validation layer (not an orchestrator). pi-teams remains the recommended orchestration. CrewAI flagged as migration target if pi-teams outgrown. Next: Enable pi-teams quality gate hooks to test whether shell-based validation suffices before considering PydanticAI.

CLAIM: Q-05 | opencode/ext-agent | 2026-05-11T02:42:00Z

Status: CLAIMED Task: Q-05 Knowledge Depot / RAG — evaluate LLM-wiki vs SQLite vs LlamaIndex vs RAG

RELEASE: Q-05 | opencode/ext-agent | 2026-05-11T02:43:00Z

Status: HANDLED (researcher's existing answer accepted) Output: llm-wiki/wiki/research/answers/q-05-knowledge-depot-rag.md (existing, researcher-authored) Summary: Researcher already wrote comprehensive Q-05 answer. Recommendation: LLM-wiki for MVP, optional SQLite indexing, avoid LlamaIndex on Pi. Matches my assessment. Marked as ANSWERED.

RELEASE: Q-07 | opencode/ext-agent | 2026-05-11T03:05:00Z

Status: HANDLED Output: llm-wiki/wiki/research/answers/q-07-toolchain-kos-contract.md Summary: KOS is external dependency, not in agent image. Documented as full contract in answer. Agents use kos.mk, never install or patch KOS. Need to verify KOS_BASE is set on Pi.

RELEASE: Q-12 | opencode/ext-agent | 2026-05-11T03:15:00Z

Status: HANDLED Output: llm-wiki/wiki/research/answers/q-12-external-agent-coordination.md Summary: Evaluated EXTERNAL-AGENT-PROMPT.md protocol effectiveness, defined hardened prompt with 6 improvements (log.md verification, stale claim takeover, config file boundaries, Forgejo credential gating, dual write requirements, interface boundary), specified collision avoidance protocol, completion semantics (HANDLED/PARTIAL/BLOCKED/RELEASED), and onboarding sequence. Existing protocol is effective; minor hardening recommended. Next: Mehdi to review 4 decisions (hardened prompt adoption, hardware constraint line, log sweep, prompt versioning). Lead (Q-00) to synthesize into architecture.

RELEASE: Q-07 | opencode/ext-agent | 2026-05-11T03:25:00Z

Status: ANSWERED Output: llm-wiki/wiki/research/answers/q-07-toolchain-kos-contract.md Summary: KOS toolchain exists at /mnt/kitchen/collab/d3-tui/toolchains/kos/ (80MB) but not in build script search paths. Fix: bind-mount to /opt/toolchains/dc/kos + set KOS_BASE env. Create kos-contract.md as canonical reference. Smoke test kos-cc before toolchain-dependent work. Next: Available for other OPEN questions.