Loader Sprint Index

These sprints translate the REPORT.md findings into implementation lanes with clear deliverables, test strategy, and definition-of-done checkpoints.

The plan was reshaped after a deeper validation pass against refs/claw-code and refs/oh-my-codex. The reshape (a) splits the most ambitious sprint, (b) reorders so the user's actual pain point (premature completion, weak follow-through) lands sooner, and (c) lands hooks alongside permissions so later sprints have a clean lifecycle to hang behavior on.

Phase 1: Runtime Foundation

• Sprint 00 — Foundation, Measurement, and Parity Harness
• Sprint 01 — Turn Engine, Tool Contract, and Capability Profiles

Phase 2: Behavioral Contract

• Sprint 02 — Definition of Done and Verify/Fix Loop

Phase 3: Safety and Workflow Discipline

• Sprint 03 — Permission Modes and Tool Lifecycle Hooks
• Sprint 04 — Mode Router, Clarify, and Plan Artifacts

Phase 4: Durability and Product Surfaces

• Sprint 05 — Session State, Memory, and Compaction
• Sprint 06 — Doctor, Explore, Status, and Tool Surface Expansion

Phase 5: Execution Policy and Runtime Simplification

• Sprint 07 — Rule-Based Permissions and Runtime Decomposition

Phase 6: Prompt Contract and Operator Ergonomics

• Sprint 08 — Prompt Builder, Runtime Phases, and Permission Operator UX

Working principles

• Each sprint must end with stronger runtime reliability, not just more features.
• Prefer behavior that improves any capable model over model-specific prompting tricks.
• Add new workflow/state surfaces only when they reduce prompt pressure or improve verification.
• No sprint is complete until its behavior is covered by automated tests or a deterministic harness.
• When adding lifecycle behavior (validation, dedup, verification, observability), prefer hooking into the tool lifecycle from Sprint 03 over patching the runtime loop directly.