ANSWERED

Open Architecture Research Queue

Purpose: answer the open planning questions before we commit to the final Pi agent setup.

Agents should claim one question, answer it in the listed answer path, then update this queue.

Status values: OPEN, CLAIMED, ANSWERED, PARTIAL, BLOCKED.

Q-00: Architecture Synthesis

Status: ANSWERED by lead, 2026-05-11T02:45:00Z Answer path: answers/q-00-architecture-synthesis.md

Question: Given all answered questions, what is the recommended minimal architecture for this Pi workcell?

Nested questions: - Should the first production shape be pi-teams only, LangGraph-driven, or a hybrid? - What should be the MVP: one always-on lead plus specialists, three parallel agents, or both? - What must be staged before agents can reliably run without getting stuck? - What is the smallest setup that preserves improvisation without losing order?

Depends on: Q-01 through Q-12.

Q-01: Pi Teams Fit

Status: ANSWERED by opencode/ext-agent, 2026-05-11T02:10:00Z Answer path: answers/q-01-pi-teams-fit.md

Question: What pi-teams features directly support the desired workcell, and where does it fall short?

Nested questions: - How do teams, teammates, task lists, messaging, plan approval, hooks, and templates work? - Can tmux panes double as the operational GUI surface well enough for MVP? - How should the lead, researcher, and builder-reviewer roles be configured? - How should tasks be created, assigned, acknowledged, and closed? - Does pi-teams provide enough shared alignment without LangGraph? - What exact teams.yaml and agent definitions should be staged?

Q-02: LangGraph Fit

Status: ANSWERED by opencode/ext-agent, 2026-05-11T02:15:00Z Answer path: answers/q-02-langgraph-fit.md

Question: Would LangGraph improve this setup enough to justify its complexity?

Nested questions: - Would LangGraph need thorough pre-staging to avoid becoming the main project? - How would state, checkpointers, retries, and graph transitions map to the T0-T7 task lifecycle? - Can LangGraph coexist with the Pi coding agent harness, or would it wrap/replace Pi? - Could LangGraph run as a lightweight scheduler while Pi agents do the coding? - What would be the smallest LangGraph proof-of-concept worth testing? - What failure modes would LangGraph introduce on a Raspberry Pi?

Q-03: Framework Comparison

Status: ANSWERED by opencode/ext-agent, 2026-05-11T03:10:00Z Answer path: answers/q-03-framework-comparison.md

Question: Compared with pi-teams, do CrewAI, AutoGen/Microsoft Agent Framework, or Pydantic AI offer a better fit?

Nested questions: - Which framework best supports open-ended improvisation plus deterministic task order? - Which framework has the lowest setup and maintenance friction? - Which framework integrates cleanly with existing Pi agents, tmux, Forgejo, and LLM-wiki? - Which framework should be rejected for MVP, and why? - Is any framework worth using only for typed outputs, planning, or evaluation rather than full orchestration?

Q-04: Work Shape / Lifecycle

Status: ANSWERED by researcher, 2026-05-11T02:20:00Z Answer path: answers/q-04-work-shape-lifecycle.md

Question: What should the per-task order of operations be?

Nested questions: - Should every issue run T0 intake, T1 review/research, T2 plan, T3 chunk/dispatch, T4 implement, T5 validate, T6 review/publish, T7 reflect/close? - What artifacts should be mandatory at each stage? - When should agents stop and ask for human decision? - How should multiple agents agree before work starts? - How should external agents participate without duplicating claims?

Q-05: Knowledge Depot / RAG

Status: ANSWERED by researcher, 2026-05-11T01:55:00Z Answer path: answers/q-05-knowledge-depot-rag.md

Question: What should the shared knowledge system be for this MVP?

Nested questions: - Is LLM-wiki enough, or should SQLite, LlamaIndex, or another RAG layer be added? - What document schema keeps answers easy for agents to retrieve and update? - What should stay in Pearl Brain versus the local LLM-wiki? - How should contradictions, source citations, and stale answers be handled? - What is the smallest useful retrieval/indexing layer that does not become a separate platform project?

Q-06: Runtime Container Shape

Status: ANSWERED by builder-reviewer, 2026-05-11T01:47:00Z Answer path: answers/q-06-runtime-container-shape.md

Question: Should the MVP use one Docker container for all Pi agents, separate containers, or pi-container-sandbox?

Nested questions: - Does one Docker container with tmux/pi-teams provide enough isolation for MVP? - Does pi-container-sandbox supersede Docker, complement it, or create nested-container complexity? - What are the exact bind mounts needed for least privilege? - How should repo, wiki, runs, logs, config, and secrets be mounted? - What are the operational controls for stop/restart/recover? - What cgroup/memory-limit issue was observed and how should it affect the plan?

Q-07: Toolchain / KOS Contract

Status: ANSWERED by opencode/ext-agent, 2026-05-11T03:05:00Z Answer path: answers/q-07-toolchain-kos-contract.md

Question: What is the best way to expose KOS/Shachi/ORCA-style toolchain capabilities to agents?

Nested questions: - What does D3-TUI actually need from KOS to make progress? - Should KOS be installed in the agent image, exposed as a documented command contract, or treated as an external service? - How should agents know what KOS is good for and when not to use it? - What docs and smoke tests are needed before agents rely on KOS? - What broke previously, and how can the new setup avoid repeating it?

Q-08: Forgejo Workflow

Status: ANSWERED by opencode/ext-agent, 2026-05-11T03:10:00Z Answer path: answers/q-08-forgejo-workflow.md

Question: How should Forgejo structure the work ledger?

Nested questions: - Should every task become a Forgejo issue before implementation? - What issue template should encode T0-T7? - How should agents comment, claim, branch, and close work? - How do we prevent agents from bind-mounting or editing more than the repo/directory they need? - How should the local wiki and Forgejo stay in sync?

Q-09: Bun / Pi Install / Model Routing

Status: ANSWERED by opencode/ext-agent, 2026-05-11T00:15:00Z Answer path: answers/q-09-bun-pi-install-model-routing.md

Question: What is the right install/runtime pattern for Pi agents on this hardware?

Nested questions: - Should Pi be installed with bun install -g @earendil-works/pi-coding-agent? - What role does Node still play for Pi CLI compatibility? - Should npm be provided by Bun only, Debian package, or not at all? - How should Kimi and MiniMax keys/models be routed between lead and specialists? - What should be baked into the image versus mounted as config?

Q-10: Validation / Smoke Testing

Status: ANSWERED by opencode/ext-agent, 2026-05-11T03:10:00Z Answer path: answers/q-10-validation-smoke-testing.md

Question: What validation setup should agents use before and after code changes?

Nested questions: - What lightweight smoke tests are realistic on the Pi itself? - Is headless Flycast useful for D3-TUI validation, or out of scope for MVP? - What should be documented as “good for” and “not good for” in the validation contract? - How should agents report inconclusive validation versus failure? - What is the smallest validation loop that prevents false confidence?

Q-11: Remote UI / Observability

Status: ANSWERED by opencode/ext-agent, 2026-05-11T03:10:00Z Answer path: answers/q-11-remote-ui-observability.md

Question: What is the best remote viewing/control surface?

Nested questions: - Should the user rely on local Mac terminal attaching to remote tmux? - Should pi-gui, pi-dashboard, or another browser UI be used? - How should Forgejo UI be exposed over Tailscale? - What logs/status files should exist for quick health checks? - What is the simplest way to see panes, team state, claims, and repo issues without SSH spelunking?

Q-12: External Agent Coordination

Status: ANSWERED by opencode/ext-agent, 2026-05-11T03:10:00Z Answer path: answers/q-12-external-agent-coordination.md

Question: How should external agents coordinate with Pi agents without confusion?

Nested questions: - What exact prompt should be sent to external agents? - What should they read first? - How should they claim one question and avoid collisions? - What should they ignore as Pi-specific runtime detail? - How should they write answers so later agents can build on them? - What should count as done, partial, blocked, or stale?

Q-13: Viable Remote Monitoring GUI

Status: ANSWERED by codex/ext-agent, 2026-05-11T03:00:00Z Answer path: answers/q-13-viable-remote-gui.md

Question: What remote monitoring GUI is actually viable for observing Pi agent activity on this hardware?

Nested questions: - Since pi-gui is local-only and pi-agent-dashboard needs Docker + bridge wiring, what actually works? - Is ttyd (tmux-in-browser) sufficient as a lightweight agent-activity viewer? - Could Tailscale Serve expose Forgejo without needing the Docker container running? - What is the smallest viable browser-accessible surface that shows agent panes, claims, and repo state? - What static HTML status page could be generated from existing state/log files? - What are the resource costs of each option on a Pi 4 already running agents? - Which option gives the best effort-to-value ratio for MVP?