An unofficial lab for the OpenCode AI agent: recipes, benchmarks, experiments, and tricks for understanding and controlling how it behaves.
The key advantage: OpenCode is a great open-source project. Unlike proprietary coding assistants, where the system prompt and context assembly are a black box, OpenCode lets you inspect everything.
You can see exactly how system prompts are built, how skill descriptions leak into context, how tool descriptions steer the model, and how AGENTS.md injects instructions on every turn. That transparency gives you the flexibility to analyze, modify, and control behavior in ways closed tools do not allow.
This repo documents those mechanisms and gives you prompts, plugins, and reference material to take advantage of that flexibility.
Not affiliated with the OpenCode project.
Also available in Spanish — all research documents have been translated.
Examines the risks of automatic AGENTS.md loading through three independent injection paths (global, project root, subdirectory). Argues that habits formed in tools with small context windows carry over to OpenCode where they cause bloat, anxiety, and latency. Documents the per-read subdirectory injection mechanism that no configuration flag can block.
A reference document that explains the three-part API call structure
(system + messages + tools) and how OpenCode assembles each part.
Covers the system prompt assembly pipeline, custom prompt resolution,
AGENTS.md injection paths, system-reminder overlays, tool definition
structure, and the instruction authority spectrum. 764 lines, the
foundation for all other research in this repo.
Six operational prompts that extract and analyze the full API call context from a running OpenCode session. Prompt 1 dumps the system, tools, and messages fields. Prompts 2-3 analyze the dump for faithfulness, refusal patterns, and contamination from training data. Prompts 4-6 extract sub-agent system prompts, mode-switching overlays, and tool definitions. Includes a quick-start workflow and a guide to moving extracted prompts between harnesses.
Replaces OpenCode's binary Plan/Build mode switch with seven user-level control flags that direct the model's cognitive mode without harness modifications. Each flag (LOCK, IDEAS, PLAN, EXPLAIN, REQUIRE, SUMMARY, EXIT) tells the model what kind of thinking to perform. Includes a ready-to-use prompt template that can be appended to any custom prompt file and is portable across harnesses.
Three experiments measuring LLM determinism on analytical tasks: single-prompt replicates (10×, σ=0.51), chained forks (3 branches, 22-33% agreement), and cross-model evaluation (4 models, 8 questions). Reports a Global Determinism Index of 0.59 (poorly deterministic) and shows that determinism depends on task granularity — convergence on coarse ranking, divergence on fine scoring. Documents the majority voting paradox (correlated bias makes consensus less reliable) and 6 mitigation strategies including multi-model orchestration.
Compares DeepSeek V4 Flash (Junior) and DeepSeek V4 Pro (Senior) as coding agents using the same custom prompt. Documents behavioral profiles (Flash: broad sweep, closure impatience, deflection under criticism; Pro: tunnel vision, security detection, multi-step tracking), a decision tree for model selection, a Flash→Pro chaining strategy, and 6 prompt rules derived from cross-model analysis.
Compares both models as context compaction models in OpenCode. Demonstrates that Flash produces better results 4× faster and 13× cheaper than Pro. Includes the "reasoning vs extraction" hypothesis, model profiles (tunnel effect investigator vs wide-spectrum explorer), escape mechanisms of both models, prompt tips per model, and the finding that identity preservation trumps evidence in both.
A manual, flat-file memory system for AI coding assistants. Uses >>/<<
operators to save and load context on demand — zero tokens until invoked.
Compares three approaches: OpenCode's AGENTS.md, an autonomous model-driven
memory, and this human-in-control alternative. Documents design principles,
workflow, operators, scopes, and a comparative analysis across 17 dimensions.
Evaluates 4 free models (DeepSeek V4 Flash Free, MiMo V2.5 Free, MiniMax M3 Free, Nemotron 3 Super Free) as analytical evaluators on 8 sequential questions over 343 KB of technical documents. Scores each model on per-question quality, global coherence, final report, operational fluency, and theoretical cost. DeepSeek leads (9.14) with zero errors and stable σ=0.35; MiMo is fastest (213s) but has format errors; MiniMax is 3.7× slower; Nemotron fails critically (4.29). Includes cross-validation with 10 replicates.
Investigates how skill descriptions automatically enter the system
prompt on every turn through the available_skills XML block
(approximately 26% of the system prompt). Includes a proof of concept
with a persona-injection skill (Grillo), evidence of real-world
degradation when skills accumulate, and a technology bias example.
Documents two mitigation approaches: a protocol-only approach (Option A)
and a plugin-based loader with on-demand loading (Option B, recommended).
A plugin that overrides OpenCode tool descriptions using plain
.txt files. Tool descriptions carry higher authority than system
prompt instructions, making this the ideal place for behavioral
rules, domain-specific constraints, and custom workflows. Also
saves tokens by shortening verbose built-in descriptions.
Used in the Skill Description Leak research as one of two mitigation options (plugin-based skill loading). Can also be combined with the Context Dump toolkit: override tool descriptions before dumping to compare how different descriptions affect the model's tool-use behavior.
Antonio Muñoz