[2026-05-11T04:30:00Z] opencode/ext-agent | INFRA | Queue injection + prompt update
Injected 6 queue items (Q-07 through Q-12) into lead inbox. Updated all three agent prompts to dispatch-driven continuous work loop: - Lead: Continuous loop — process one inbox item at a time, dispatch to researcher/builder-reviewer, write answers, update queue. - Researcher: Dispatch-driven — wait for lead dispatch, no self-claiming. - Builder-reviewer: Dispatch-driven — wait for lead dispatch, plan approval before implementation.
Restarting container to pick up new prompts and begin processing queue items. ## [2026-05-11T18:15:00Z] opencode/ext-agent | SYNTHESIS | LangGraph as Turn-Based Strategy Game Surface
Status: WRITTEN Output:
/workcell/llm-wiki/wiki/architecture/langgraph-game-surface.md
Source: MjF + sandshrew conversation exploring LangGraph core
mechanisms, pre-staged vs emergent behavior, tool calls without
LangChain, game surface concept, RTS engine evaluation, and
cross-platform console client architecture. Summary: LangGraph’s
primitives map directly to turn-based strategy game mechanics. Nodes =
hexes, edges = paths, state = unit positions, super-steps = turns.
Lightweight Python renderer (pygame) recommended over RTS engines.
Three-console client architecture (RG40XXV/Wii/PS2) connecting to Pi
LangGraph backend. Orthogonal to Q-02 — this is a game exploration, not
a workcell orchestrator. ## [2026-05-11T18:25:00Z] opencode/ext-agent |
SYNTHESIS | LangGraph Gameplay Modes — Auto-Run vs Control + Per-Unit
Toggles
Status: WRITTEN Output:
/workcell/llm-wiki/wiki/architecture/langgraph-gameplay-modes.md
Summary: Three execution layers: control mode (interrupt() pauses per
human unit), auto-run mode (no interrupts, AI loops), per-unit/lane
toggles (auto_pilot flag on each unit, checked by conditional edge).
Toggle is live — flip at any interrupt point, next cycle picks it up.
Turn cycle: Planning → Execution (Send fan-out) → Resolution →
End-of-Turn gate. Per-lane grouping with override semantics. Six edge
cases identified. Four decisions needed from Mehdi. ##
[2026-05-11T18:35:00Z] opencode/ext-agent | SYNTHESIS | Hex/Tile Node
Mapping — Concept Exploration
Status: WRITTEN Output:
/workcell/llm-wiki/wiki/architecture/langgraph-hex-node-mapping.md
Summary: Explores modeling map hexes as individual LangGraph nodes.
Contrasts with traditional passive terrain data. Observes that Pregel
execution activates only entered hexes (not all nodes), making per-turn
cost scale with unit movement rather than map size. Notes hybrid
spectrum — active locations as nodes, open terrain as state data. Poses
four open questions about compile-time limits, dynamic node generation,
renderer/graph separation, and LangGraph internal limits. ##
[2026-05-11T18:45:00Z] opencode/ext-agent | SYNTHESIS | LangGraph Node
Anatomy — Full Catalog
Status: WRITTEN Output:
/workcell/llm-wiki/wiki/architecture/langgraph-node-anatomy.md
Summary: Catalog of everything a LangGraph node can contain, organized
across five dimensions: INPUT (state, config, runtime), PROCESS
(deterministic logic, side effects, control flow, limitations), OUTPUT
(state update, return variants, side-stream), REFERENCES (internal
graph-aware, external runtime-injected, node-local), CONSTRAINTS
(signature, return type, write scope, serialization, caching). Not
mapped to game concepts — raw anatomy for later conceptual mapping. ##
[2026-05-11T18:55:00Z] opencode/ext-agent | SYNTHESIS | LangGraph Tool
Execution Tradeoffs — LangChain vs External Harness
Status: WRITTEN Output:
/workcell/llm-wiki/wiki/architecture/langgraph-tool-execution-tradeoffs.md
Summary: Trade-off exploration between LangChain tools (full visibility
in state, checkpointable, interruptible, but pre-staging every tool
schema) vs external agent harnesses like Pi/Hermes (adaptive,
lightweight, no pre-staging, but internal tool calls opaque to LangGraph
state). Three bridging patterns observed: Black Box (agent runs
autonomously), Full Visibility (LangChain ToolNode), Hybrid (LangGraph
routes + provides situation, Agent decides + reports high-level
outcomes). Notes LangGraph is indifferent to internal implementation —
LangChain tools are NOT the native convention. Open questions about tool
visibility necessity, subprocess overhead, streamed events, and
interrupt needs. ## [2026-05-11T19:10:00Z] opencode/ext-agent |
SYNTHESIS | Unit-Node Interaction Model — Full Concept
Status: WRITTEN Output:
/workcell/llm-wiki/wiki/architecture/langgraph-unit-node-interaction-model.md
Summary: Full observational mapping of MjF’s proposed model. Units =
state cases with roles, configs, access lists, node overrides. Nodes =
task spaces (research/design/dev/review) with output locale pointers.
Access gating via allow/block lists. Per-node config overrides
(model+harness switching). Trailing context via state persistence.
Passive intel harvesting: harvester units prospect, flag findings as
named modular intel items (“pi torch”), auto-suggest routing to target
nodes. Human gates acceptance via interrupt(). Five open questions
posed. Confirmed: maps cleanly to LangGraph primitives — state is the
unit’s backpack, nodes are stateless functions over unit state. ##
[2026-05-11T18:55:00Z] opencode/ext-agent | INFRA | Forgejo Repo Created
+ Layered Questions Staged
Status: COMPLETE Output: - Repo:
/home/mehdifarah/git/langgraph-game-surface.git on Pi
(bare) - Clone:
ssh://mehdifarah@100.120.38.37:/home/mehdifarah/git/langgraph-game-surface.git
- Wiki:
/workcell/llm-wiki/wiki/architecture/langgraph-layered-questions.md
- Forgejo: http://100.120.38.37:3001/ (needs configuration
— app.ini absent, web install pending) Summary: Created bare git repo
for langgraph-game-surface project on Pi. Seeded with README + 27
layered open questions organized across three layers: Pi
(hardware/runtime/harnesses — 7 Qs), Game (assets/UX/engine — 10 Qs),
Membrane (wiring between Pi and Game — 10 Qs). Questions trace back to
wiki architecture pages as sources. Forgejo container is running but not
configured (no app.ini) — needs web install step before it can serve the
repo via web UI. Currently accessible via SSH git:// only. ##
[2026-05-11T19:00:00Z] opencode/ext-agent | NOTE | Forgejo Web UI —
Deferred
Forgejo container is running on Pi (port 3001) but app.ini is absent — the instance was started but never configured via web install. The web UI at http://100.120.38.37:3001/ shows the install/setup page. Bare git repo works fine over SSH at ssh://mehdifarah@100.120.38.37:/home/mehdifarah/git/langgraph-game-surface.git. Forgejo web configuration deferred to a later session — not blocking current work. ## [2026-05-11T19:20:00Z] opencode/ext-agent | RESEARCH | Engine Evaluation — 7 Engines Surveyed
Status: WRITTEN Output: - Wiki:
/workcell/llm-wiki/wiki/architecture/langgraph-engine-evaluation.md
- Repo: docs/ENGINE_EVALUATION.md in
langgraph-game-surface.git Summary: Surveyed 7 engines for the rendering
layer. Wargame Engine (maximinus/wargame, Python+Pygame, 2024)
recommended as best fit — its node-tree viewport, scene system,
messaging architecture, and board wargame domain align architecturally
with LangGraph’s graph model. Fallback: raw pygame. Fabula unfindable
(likely doesn’t exist on GitHub). Godot/Panda3D/OpenRA rejected — wrong
runtime, wrong domain. Sequtus worth skimming for Pygame RTS patterns.
Forgejo web UI configuration deferred to later session.