dogma: Instructions file for Codex

AGENTS.md

Guidance for AI coding agents working in this repository.

Guiding Constraints

These constraints govern all agent behavior. They derive from three core axioms in MANIFESTO.md:

1. Endogenous-First — scaffold from existing system knowledge and external best practices
2. Algorithms Before Tokens — prefer deterministic, encoded solutions over interactive token burn
3. Local Compute-First — minimize token usage; run locally whenever possible (LCF as structural enabler — see MANIFESTO.md §3)

Encoding Inheritance Chain: Values flow through MANIFESTO.md (foundational axioms) → AGENTS.md (operational constraints) → role files (.agent.md; VS Code: Custom Agents) → SKILL.md files (reusable tactical knowledge) → session prompts (enacted behavior). Each layer is a re-encoding of the layer above it. Agents must minimise lossy re-encoding: prefer direct quotation or explicit citation over paraphrase when invoking a foundational principle. Cross-reference density (back-references to MANIFESTO.md in your output) is a proxy for encoding fidelity. Low density signals likely drift. See docs/research/values-encoding.md for the cross-sectoral evidence base. AGENTS.md is the central authority for all operational constraints; subdirectory redirection notices point back here for global governance.

Session-Start Encoding Checkpoint: At the start of every session, the first sentence of ## Session Start in the scratchpad must name the governing axiom and one primary endogenous source. See docs/guides/session-management.md → Session-Start Encoding Checkpoint for format and examples. The agent fleet is the pressurizing medium — it gives each substrate coherent form but does not own the membrane or the bucket. Agents are tools that shape how values flow through the system, but they neither create nor control the values themselves.