OpenExec V1 Simplification Pass
To ensure OpenExec remains a minimal, deterministic, and maintainable AI coding runtime, the following "7 Deletions" must be executed during the v1 stabilization phase.
1. Delete Dual Runtime Behavior
Problem: Separate engines for chat, execution, and planning create duplicate state, prompts, and routing logic. Solution: Move to one canonical flow: Session → Task → Run.
- Chat is strictly the entrypoint.
- Execution is strictly a Run.
- No parallel logic trees or "brain" duplication.
2. Delete Legacy JSON State as Live Truth
Problem: Relying on stories.json or tasks.json alongside SQLite creates "archaeology" instead of a single runtime truth.
Solution: SQLite is the only authoritative state.
- JSON/YAML exist only as input definitions, exports, or debug artifacts.
- If recovery depends on a JSON file, that code must be removed.
3. Delete Prompt-Driven Workflow Control
Problem: If prompts decide the next stage or whether to test, the prompt is secretly the workflow engine. Solution: Transitions, retries, and routing belong in the Blueprint Engine.
- LLM responsibility is limited to: Implementation, Diagnosis, Summarization, and Review Reasoning.
4. Delete Flat Tool Universes
Problem: Exposing all tools by default causes model confusion and increases leakage risk.
Solution: Enforce Curated Toolsets (e.g., repo_readonly, coding_backend).
- Each run receives exactly one default toolset with tightly controlled access.
5. Delete Arbitrary Shell as a First-Class Primitive
Problem: shell("do whatever") destroys determinism, replayability, and policy enforcement.
Solution: Prioritize Typed Deterministic Actions (e.g., apply_patch, run_tests).
- Arbitrary shell exists only as a high-risk, approval-required capability.
6. Delete Premature Multi-Agent/Distributed Ideas
Problem: Swarms, remote workers, and complex delegation layers are second-system traps for v1. Solution: Stick to One Runtime, One Engine, One Active Run.
- Focus on runtime discipline over "agentness."
7. Delete Overlapping Orchestration Models
Problem: Coexisting concepts (legacy phases vs. new blueprints) make the system impossible to reason about. Solution: The Blueprint Engine is the only execution model.
- All other flows (e.g., analytical tasks) must become artifacts or helper subsystems for the blueprint.
Descoping Priorities (Postpone)
- A. Big UI Ambitions: Focus on CLI and terminal timeline first; dashboard is secondary.
- B. Broad MCP Ambitions: Stick to a local tool registry before building a "Toolshed" platform.
- C. Hyper-generic Abstraction: Normalize I/O and streaming; don't try to hide every model difference.
Hard V1 Rule
If a feature does not clearly improve determinism, local control, state clarity, tool safety, or replayability, it must be postponed.
Summary: OpenExec wins by being a small core with typed actions and explicit workflows, not by having more "magic" than its competitors.