Triphasic Execution Framework

What problem it addresses

When an AI agent handles complex work, the failure mode is often not a single tool call, but process drift: it skips constraint analysis, plans without confirmation, executes multiple steps without review, retries the same failing approach, or loses progress after an interruption. triphasic-execution turns these constraints into an explicit execution framework built around execute → review → advance as the smallest unit of each interaction.

How the skill works

The skill uses a progressive Markdown layout: SKILL.md remains the entry point, while detailed rules are split into references/*.md and loaded on demand. Its core workflow includes:
- Semantic decomposition: extract 5W2H details and constraint strength before planning.
- Task planning: emit a confirmable plan, then persist state with task_progress.py init.
- Triphasic loop: each step must be executed, reviewed, advanced, and then progress-updated.
- Failure control: limit each step to 3 retries and 3 idle turns, forcing a new approach or user input when limits are reached.
- Post-task capture: record problems, risks, and lessons with problem_logger.py, then merge pending entries into LESSONS_REGISTER.md.

It fits multi-step agent workflows that need progress tracking and resumability, but is overkill for simple questions, one-step commands, or pure information lookup. It also relies on scripts and hooks for enforcement, so the surrounding runtime must support those checks.