geode-context

name	geode-context
visibility	public
triggers	geode, harness, agent, runtime, release, packaging, docs, plugin, architecture
description	GEODE runtime architecture and repo-shape context. Use for architecture, packaging, release, docs, CI, AgenticLoop, async runtime, and plugin boundary questions.

GEODE Context

Version-agnostic runtime context. Use geode version / geode about for the installed version and runtime facts; never trust numbers remembered from this skill.

Current Shape

• GEODE is a general-purpose autonomous-agent harness whose runtime is AgenticLoop(while tool_use).
• The agent loop, sub-agents, plans, and batches are all instances of the same tool-use loop.
• The runtime is domain-agnostic. Shipped packages:
  • core/ — agent loop, tools, MCP, memory, hooks, wiring, CLI, server, gateway.
  • evals/ — Petri audits, benchmarks, seed generation, and GEO measurement.
  • evolve/ — scaffold search and Crucible experiment supervision.
• site/ is the public Next.js docs/site, not a Python package.
• Dependency direction is evolve -> evals -> core; reverse imports are forbidden.

Runtime Boundaries

• Canonical execution path is async-first: AgenticLoop.arun(), tool aexecute(), async provider clients.
• core/agent/loop/agent_loop.py is the implementation. DEFAULT_MAX_ROUNDS = 0 means no round cap. Model completion and runtime guard exits are distinct TerminationReason values in core/agent/loop/models.py; a diagnostic or budget exit is not task success.
• Sub-agents: core/agent/sub_agent.py (SubAgentManager) — max depth 1 (no recursion), session-wide cap 15, global Lane concurrency 50 (core/wiring/container.py).
• Headless mode policy lives in core/server/supervised/services.py: run_bash and delegation remain denied; DAEMON desktop control requires explicit gateway.allow_computer_use, while SCHEDULER keeps the desktop denial. The active tool plan and profile can further restrict access.

Prompt And Context Injection

• core/agent/system_prompt.py builds the base prompt; core/agent/loop/_context.py composes skill and session context. core/llm/prompt_assembler.py contains shared formatting helpers, not another assembler.
• PROMPT_CACHE_BOUNDARY is the opening <dynamic_context> tag. Stable instructions precede it; per-turn context stays inside the closed envelope. core/agent/loop/agent_loop.py emits PROMPT_ASSEMBLED after each per-round rebuild.
• Skill metadata is injected as <available_skills> (metadata only); full skill bodies load on demand via the use_skill tool.
• Do not inject old Game IP pipeline facts, fixed DAG claims, analyst/evaluator topology, or confidence-threshold loop claims into the system prompt.

Release Pipeline

• Functional commits update CHANGELOG.md.
• Quality gates: ruff check / ruff format --check (core, evals, evolve, tests, scripts), mypy core/ evals/ evolve/, lint-imports, pytest -m "not live", geode version smoke.
• Publishing is manual/approval-gated, not automatic on every main push.

Guardrails

• Do not restore the retired top-level plugins, geode_product, or core.self_improving packages. Evaluation and evolution stay explicit outer consumers of core.
• Do not describe GEODE as a fixed Plan-and-Execute DAG, fixed StateGraph, or confidence-threshold pipeline.
• Do not cite LangSmith as current observability; GEODE uses its own hooks, audit diagnostics, run logs, and site/docs gates.
• Do not hard-code deprecated model names. Use the provider registries and token/cost tables.