GPT-6 Astra Agent Config: Standing Rules for AGENTS.md

Working note on GPT-6 Astra agent configuration. Core standing rules to remove old braking behavior and replace unconditional stop lines with action-class conditions: R3-R6 (AGENTS.md): replace unconditional "ask before / always test / stop after first / read all docs" lines with conditional pointers and action classes. Test wording: "Run tests appropriate to the change and complete required checks." R9 (prompt / AGENTS.md): DONE MEANS / NOT DONE MEANS / ROUTE / STOP FOR ME ONLY IF. Done lines must be script-checkable. R10 (AGENTS.md): AUTHORIZED FOR THIS TASK / NOT AUTHORIZED lists. Example: "The local tests use disposable fixtures and have no production access. Run them without asking for approval at each step." R11 (AGENTS.md): guide prompts so user instructions take precedence over skills; on any pause, name and link the exact SKILL.md file read. R12 (config.toml): [[hooks.Stop]] -> [[hooks.Stop.hooks]] type="command" command="python3 .codex/hooks/contract.py"; JSON on stdout {"decision":"block","reason":"..."}; guard on stop_hook_active. R13 (config.toml / .codex/agents): approval_policy="on-request"; [[hooks.PermissionRequest]] returning decision.behavior="allow" for the test runner; verifier agent TOML with required name, description, developer_instructions, plus sandbox_mode="read-only". R14 (config.toml): model="gpt-6-astra", model_reasoning_effort="low"; [agents] default_subagent_model="gpt-5.6-luna"; [features.context_management] experimental_mode=true. Acceptance test before trusting the setup: confirm the instruction files actually loaded match what you expect, grep AGENTS.md and SKILL.md for leftover unconditional "ask before / always / stop after" lines, check the skills catalog fits its context budget, verify every unattended task has a script-checkable DONE list, confirm the Stop hook exists and fires at most once per turn, and make sure any pause names the file and rule that caused it.