Skip to content

Agent Instruction Changelog

Every change to agent behavior — AGENTS.md, .claude/agents/*.md, .claude/skills/, .github/copilot-instructions.md, or the policy data under _data/agents/ — gets an entry: what changed, why (the observation that motivated it), and the measured outcome. One variable at a time. This is the GH-600 Domain 4 discipline of treating instructions like code; the format is taught in Chapter IV — The Oracle Rubric.

2026-07-01

AGENTS.md

  • Added: the "🚫 Forbidden Actions (the Warden Pact)" section — an explicit list of actions no agent may take regardless of instructions, with the decline-comment-label-stop protocol.
  • Reason: the fleet's boundaries lived implicitly across nine workflow header comments and agent definitions; nothing stated them in the one file every agent reads first. GH-600 Domain 6 implementation surfaced the gap.
  • Outcome: TBD — first measure at the next weekly agent-audit.yml fleet audit (does the auditor find zero pact/reality drift?).

_data/agents/ (new policy surface)

  • Added: registry.yml (fleet roster: lane, kill switch, status, review date per agent) and autonomy-matrix.yml (recurring action → L0–L4 autonomy level + the guardrails that justify it).
  • Reason: lifecycle management (D5) and risk-based autonomy (D6) need a machine-readable source of truth, not folklore; the weekly audit now has a contract to check the fleet against.
  • Outcome: TBD — measured by audit findings over the next quarter (review_date: 2026-10-01 on every roster row).