task-tree-core-statelisted
Install: claude install-skill guess-guess-who-i-am/llm-task-tree
# Task Tree Core State
Treat the live tree as working memory, not an audit log. Preserve enough state to choose and verify the next action; move recoverable detail to evidence files.
## Core-State Gate
Use `GraphState.Next` and that node's `NextIdea` as the query. For every candidate statement, apply this deletion test:
> If removed, could the next Agent choose a different method, repeat a disproved action, violate a live constraint, miss an unresolved risk, or lose the completion test?
Classify privately; do not write these labels into the tree:
- **Retain**: current problem/method, user decision, active constraint, unresolved risk, up to three decision-changing results, one executable `NextIdea`, and short evidence pointers.
- **Move**: useful but recoverable process, stable architecture/background, raw output, file inventories, cases, and implementation detail. Put these in the appropriate tree, ADR, or `scripts/steps/`; keep one pointer only when needed.
- **Drop**: duplicate, superseded, resolved, timestamped, or process-only text already recoverable from `versions/`.
Retained prose defaults to concise Chinese. Familiar terms such as `LLM`, `token`, `API`, exact names, IDs, paths, and URLs may remain. Move code, JSON, commands, formulas, raw data, logs, and complex English domain terminology to evidence files.
## Workflow
1. Read focus with `task_tree_focus`; read the Next node and only the dependencies needed for its `NextIdea`.
2. Measure the affected tree wit