Adaptive Prompt Policy

# Adaptive Prompt Policy — Context-Adaptive Spawn Tuning

**Purpose:** Auto-tailor every spawn prompt to the current **project** and **session** context, and self-reinforce within the session from observed outcomes — so inter-agent instructions get sharper as a session progresses, without any durable global rewrite.
**Read when:** Composing a spawn prompt at an EXECUTE step, or deciding how directives adapt to project/session signals.

> **Scope is the safety model.** This policy operates **only within the current session + project**. It is **ephemeral** (resets at the session/project boundary) and **reversible** (every adjustment is per-spawn; nothing irreversible happens). Because no durable global file is written, **no approval gate is required** — this runs automatically in all modes. Durable, cross-project template rewrites are explicitly **out of scope** here; that path stays gated (offline `tune` → Darwin promotion → Guardian commit, see §6).

> **Honest mechanism.** This is **evidence-accumulating, case-based adaptation** — not neural RL. The hub cannot train weights mid-session. "Reinforcement" means: a journaled within-session record of `context-features → directive-choice → outcome`, consulted to bias the next spawn's directive selection. Bounded, **corrective (bidirectional)** heuristics over a vetted directive library, never free-form prompt invention — every adjustment maps to an existing structured directive field (envelope / effort / tool-use / thinking / which references), never raw prepended text.

---

## 1. The three layers