Continual learning proposal for Codex

Use this only when the user explicitly asks to mine supplied task history or when an operator-enabled trusted project workflow invokes it.

Workflow

1. Verify the transcript is explicitly authorized, belongs to the current workspace/task scope, and excludes secrets and unrelated private conversations.
2. Delegate the transcript review to one bounded native memory curator with an explicit proposal-only briefing. The coordinator supplies the exact input path and requests a proposal-only result.
3. Ask the curator to identify durable candidate rules, cite their source lines or hashes, deduplicate them against current instructions, and distinguish user authority from temporary task wording.
4. Return a reviewable proposal and receipt. Do not edit AGENTS.md, global memory, repository instructions, hooks, or runtime state unless the user separately authorizes that exact write.
5. If the transcript or native delegate is unavailable, return not-run and leave all outputs unchanged.

Automatic Stop-hook triggering remains unavailable. The observed native proof reviewed one explicitly authorized local transcript, produced a hashed proposal, and performed no repository, instruction, memory, hook, or global-state write.