Repository navigation
Single source of truth for AI rules across agents, AGENTS.md, extensions, skills, Cratis AI corpus and other harnesses #30
Description
woksin (2026-10-23, via SPEC): "setup and rules should live in ONE place, not copied into agent definitions, prompts, skills or per-harness files. Please research this before choosing."
Rules are copied today into agent definitions (agent/agents/*.md), the global and profile AGENTS.md, extensions and guards, skills, the Cratis AI corpus (.cratis/ai, installed with cratis ai install/update) and other harnesses' files (Claude Code, Copilot, Codex). The copies drift. Agent types then behave differently under the same policy, as in #29.
Research questions
- Where do rules drift today? An inventory of the duplicates, with examples of drift.
- Which mechanism gives one source of truth with per-repo and per-task overrides that every agent type honours equally? Options:
- generated agent files;
- runtime injection by an extension;
- enforcement by hooks or tools instead of prose;
- layered config.
- How does it fit the Cratis AI corpus distribution and other harnesses?
Deliverable
A decision document, the method documented in the pi-setup README, and the migration, in slices. #29 (the pi-phase execution policy) is the first concrete case.
Metadata
Metadata
Assignees
Labels
No labels
Activity
woksin commentedon Oct 23, 2026
Research done; the decision is proposed and awaits woksin's open questions.
noContextFiles, so every rule reaches them only through copies in the agent files: 8 agents × 8 profile directories = 64 copies. Those copies already disagree on priority-high, the scripting rule (missing from the default profile's implementer) and the execution exceptions.policyextension delivers the subagent sections at runtime. Mechanical rules go into tools and guards (pi-phase decides admission). Agent files keep only their role. Profiles become overlays. Cratis rules stay in the Cratis AI corpus and are referenced, not copied.