RDL extension: Context Compilation Directives for agent dispatch

[zenprocess]

Background

Building a context compiler for AI coding agent dispatch surfaced a set of concepts that should be standardized in the RDL format:

1. Compression tiers — files have different value at different compression levels (full content, structural skeleton, NL summary, manifest entry, omitted). The tier assignment is task-dependent.

2. Scope tags — instruction sections tagged with domain scopes (backend, frontend, api, testing) enable per-task filtering. Format: HTML comments in markdown.

3. Tier utility matrices — different file types have different information density at each tier. A config file's skeleton is worthless but its full content is critical. A type file's skeleton carries 90% of the information.

4. Submodular value — context items aren't independent. Adding file A reduces the marginal value of file B if B re-exports A's types. The optimization is submodular, not knapsack.

Proposed RDL Extensions

1. Scope directive

# scope: backend, database
RULE:no-raw-sql
  EFFECT:forbid
  ...

Rules tagged with scope are only included when the task's scope overlaps.

2. Tier directive

# tier: skeleton
# tier-utility: 0.9
RULE:type-definitions

Advisory hint for context compilers about how to compress this rule under budget pressure.

3. Budget pressure behavior

# pressure: drop-at-moderate
# pressure: keep-always

Explicit control over what gets dropped first when the context budget is tight.

Why standardize

Multiple tools consume RDL: dispatch agents, IDE plugins, MCP servers, CI checks. Standardizing compression and scoping directives means a single rule file works across all consumers with consistent behavior.