agent-engineering

Agent Engineering Governance

Index of governance and engineering policy documents. Each agent loads the relevant subset on demand via its ## Resources section.

Core contracts

• PART-SPEC.md — P.A.R.T. specification and mandatory section order (Prompt → Archive → Resources → Tools) for every agent file.
• PROMPT-BEHAVIOR-CONTRACT.md — Behavioral invariants complementing P.A.R.T structural rules (evidence discipline, follow-through, ABSTAIN, escalation).
• RELIABILITY-GATES.md — Verification gate requirements (build, tests, lint, approval) shared across all agents.

Clarification and routing

• CLARIFICATION-POLICY.md — When to invoke vscode/askQuestions vs. return NEEDS_INPUT to the conductor.
• TOOL-ROUTING.md — Routing rules for external tools (fetch, githubRepo, MCP) and local-first precedence.

Scoring and consolidation

• SCORING-SPEC.md — Quantitative scoring reference for reviewers and eval harnesses.
• MIGRATION-CORE-FIRST.md — Shared implementation backbone pattern and consolidation exit criteria.

Process

• ADR-PROCESS.md — Architecture Decision Record workflow: when to author an ADR, lifecycle states, and review expectations.
• ADR-TEMPLATE.md — Canonical ADR template for recording context, decision, and consequences.

Portability & IDE Support

The repository contains adaptations of the ControlFlow system for non-VS Code environments and IDE integrations:

• Codex CLI: The ControlFlow for Codex plugin now lives in its own repository at github.com/Smithbox-ai/ControlFlowCodex.
• Claude Code: See plugins/controlflow-claude-code/README.md for the native plugin skills, agents, and local development lifecycle.
• Cursor IDE: See CURSOR-SUPPORT.md for rules, skills, subagents, and plugins/controlflow-cursor/. Optional CI patterns: CURSOR-SDK.md. Cursor support does not alter VS Code tool grants.

Runtime capabilities (shipped in the ControlFlow revision program)

• MODEL-ROUTING.md — Logical model-role indirection backed by governance/model-routing.json; actively resolved at runtime by Orchestrator and Planner via agent/runSubagent.
• OBSERVABILITY.md — UUIDv4 trace_id propagation across delegation and report schemas; NDJSON event sink convention under plans/artifacts/observability/ (one file per task).
• MEMORY-ARCHITECTURE.md — Three-layer memory model: session (volatile) / task-episodic (plans/artifacts/<task-slug>/) / repo-persistent (NOTES.md).
• AGENT-AS-TOOL.md — MCP forward-compatible subagent input contract (scope, context_refs, trace_id, iteration_index) for future native tool surfacing.