task-daglisted

# Task DAG Structure a hard goal as a directed graph of branches. **Deterministic trunk, AI only on the leaves that need it.** ## Flow 1. **plan** — decompose the goal into N branches (fan-out, e.g. `2,2,2` → 1 + 2 + 4 + 8 = 15 nodes). 2. **reflex** — every branch is classified *before* any model call: - `block` — destructive/risky (delete, wipe, drop, shutdown, format) → refuse without spending tokens - `sql` — answerable from a database - `skill` — a known procedure/reference exists - `ui` — a screen action (open/click/type) - `fastpath` — static answer - `ai` — only now call a model 3. **execute** — deterministic handler or the cheapest capable model. 4. **verify** — every branch must return **evidence** (a result, a row count, a screenshot hash). No evidence → retry, then drop the branch. 5. **merge** — join only verified branches. 6. **memory** — append findings to a log so the next run is cheaper. ## Why the reflex layer first - Destructive input (`rm -rf /`, `DROP TABLE`, `shutdown`) is refused at ~50 ms with **zero** model cost. - Most branches are SQL/skill/fastpath → the AI bill drops by an order of magnitude. ## Fan-out template ``` python task_dag.py "<goal>" --levels 2,2,2 # 15 nodes, simulated python task_dag.py "<goal>" --levels 2,2,2 --real # real workers python task_dag.py "<goal>" --levels 3,2 --proliferate 3 # same sub-task on 3 channels, verifier picks the proven one ``` ## Rules - Safety check **before** AI. Always.