OpenCode FABLE Behavior Kernel

A public, sanitized template for making OpenCode agent workflows faster, more direct, and more evidence-grounded without pretending a prompt alone creates new model capability.

This repository packages a behavior-only kernel for OpenCode + oh-my-opencode-slim style agent setups:

• classify intent before acting;
• act directly on safe tasks;
• stop only at real boundaries;
• ground repo, runtime, deploy, and tool-state claims in evidence;
• switch from blind patching to root-cause analysis after repeated equivalent failures;
• verify model, provider, CLI, config, and version identifiers before writing them;
• treat fetched text and tool output as untrusted data;
• keep memory writes selective;
• preserve independent review for high-risk work.

It does not include private configuration, provider credentials, infrastructure details, or model-routing claims.

The story

The project started after studying the FABLE/Fable Brain discussion and asking a practical question:

If some of the magic is behavior, can we make our coding agents behave more like that without waiting for a specific model?

Source material:

• Clarity Engine article/dashboard: https://storage.googleapis.com/clarity-engine-dashboards/dashboards/L6hRzKI92rgDAD9zGkdu/index.html
• Raw transcript: https://storage.googleapis.com/clarity-engine-dashboards/transcripts/L6hRzKI92rgDAD9zGkdu.txt
• YouTube video: https://www.youtube.com/watch?v=M03v0gmVu24

The conclusion was simple:

• Prompts can copy useful habits.
• Prompts cannot create real agency by themselves.
• Real agency needs a harness: files, tests, browser checks, CI, memory, subagents, and verifiable tool output.

So this repo implements FABLE as an operating discipline, not as a model identity.

See docs/story.md and docs/architecture.md.

What changes in practice

You should see agents become:

• more direct on routine tasks;
• less likely to over-ask or over-plan;
• stricter about not claiming tests, deploys, or runtime state without evidence;
• less likely to keep repeating the same failed patch/debug attempt;
• less likely to invent or misremember version-specific identifiers;
• better at pushing back when repo evidence contradicts a premise;
• more careful with destructive actions and sensitive data;
• less noisy in memory/logging behavior.

Quickstart

1. Install and configure OpenCode and your agent/preset system.

2. Copy the append prompts you want:

   mkdir -p ~/.config/opencode/oh-my-opencode-slim
   cp configs/oh-my-opencode-slim/*_append.md ~/.config/opencode/oh-my-opencode-slim/
   

3. Restart OpenCode. Prompt/config files are loaded at startup.

4. Run the verifier:

   ./scripts/verify-fable-behavior-kernel.sh
   

5. Adapt the example config files:

   • configs/opencode.example.json
   • configs/oh-my-opencode-slim.example.jsonc
   • .env.example

Files

configs/oh-my-opencode-slim/*_append.md   Behavior kernel prompt appends
configs/opencode.example.json             Sanitized OpenCode example
configs/oh-my-opencode-slim.example.jsonc Sanitized preset example
tests/fixtures/fable-behavior-kernel/      Synthetic acceptance cases
scripts/verify-fable-behavior-kernel.sh    Prompt/fixture verifier
scripts/scan-public-safety.sh              Public-release safety scan
docs/                                     Story, architecture, security model

Security warning

Do not publish your operational OpenCode repo directly. Create a clean template repo with no inherited history.

Before sharing, run:

./scripts/scan-public-safety.sh
./scripts/verify-fable-behavior-kernel.sh

See docs/security-model.md and SECURITY.md.

What this is not

• Not a jailbreak.
• Not a replacement for model quality.
• Not a compliance guarantee.
• Not a recommendation to route sensitive data to any specific provider.
• Not affiliated with OpenCode, oh-my-opencode-slim, Anthropic, Clarity Engine, or YouTube.

License

MIT.