Skip to content

AGENTS.md

Latest commit

 

History

History
26 lines (17 loc) · 2.31 KB

AGENTS.md

File metadata and controls

26 lines (17 loc) · 2.31 KB
Raw

Repository operating contract

Boundary

This repository is a research incubator for the provisional FLR-H application-runtime profile. Keep sourced mechanisms, local synthesis, executable evidence, and efficacy claims distinct. Jaebaeman, planning/dispatch mythology, a universal soup graph, ambient effects inside pure kernels, and exactly-once claims for external systems are out of scope.

Always-on governance rules

  • GOV1 — Label material assertions HYPOTHESIS, PROPOSED, MEASURED, FALSIFIED, or ACCEPTED. Only a named command, fixture, environment, and receipt can support MEASURED; prose agreement, import success, model verdicts, and trace spans cannot.
  • GOV2 — Engine promotion remains deferred. Do not build a broad runtime framework until every gate in docs/adr/0001-defer-engine-verdict.md and spec/engine-decision.v1.json is met.
  • GOV3 — Visibility, licensing, package publication, and releases are separate authorities. Require explicit user direction for each action; never infer one from another.
  • GOV4 — Preserve user changes, keep commits scoped, and never stage unrelated files.

Rule routing and context

For material work, select the smallest matching route or routes in agent-rules/routes.v1.json. Load only their declared rule sources plus the directly touched contract, fixture, validator, or receipt. When routing is ambiguous, load the union in canonical order and deduplicate by rule ID. Manifest closures and inherited digest pins are integrity evidence, not preload lists.

Prompt context is quality-first. Never impose a hard prompt-token limit, truncate required instructions, omit a required rule, or fail work solely because of prompt-token count. Context size may be recorded as an observation. Runtime queue, wait, deadline, rule, attempt, and effect budgets remain hard safety invariants and are unrelated to prompt-context observation.

Change protocol

  • Load agent-rules/semantic-invariants.md for semantic, runtime, conformance, fault, or evidence changes.
  • Add a failing conformance or fault fixture before repairing a semantic bug.
  • Validate machine-readable contracts and JSON before committing.
  • Update claim/status documents when implementation evidence changes.
  • Record exact test commands and outcomes; distinguish simulated adapters from real platform evidence.