Two operator primitives, both queued (never synchronous — CPO hard constraint), all mutations in the cron handler after re-validating eligibility UNDER the generation lock: - Replay (Stage_Replay + on_replay_job): one governed chat call from the latest edited input fork. options_from_body() is the exact inverse of build_body()+apply_reasoning_budget() (headroom subtracted back out; reasoning pinned explicitly so global-setting drift can't reshape a replay). Goes through send_chat_completion -> cost governor reserves against the SAME run row (guardrail 4), retry/empty-guard machinery, B1 recorder. Fresh output lands in the stage snapshot; downstream stale-marked (guardrail 3); human_modified set at execution entry (guardrail 2). - Rebuild (on_rebuild_job): demote target+downstream to pending, re-enter Article_Worker with the idea from the stored seed; upstream done stages reuse (never re-charged), demoted stages rebuild prompts from CURRENT snapshots — how an upstream edit propagates. Decisions recorded while rebuilding a human-modified item are flagged. - Rerun_Eligibility: published/future/private = frozen (guardrail 5, server-side); executing runs blocked; editable stages = writer chat stages (outline/draft/polish); chain-order downstream sets. - Run_State::reopen(): atomic terminal->running gate, ceiling re-snapshotted from the current setting, accumulated spend NEVER reset. - Publisher: re-run output refreshes the existing UNPUBLISHED post (status preserved — re-runs never publish); published posts untouched. - Article_Worker persists the exact idea payload as a seed (INSERT-once) so rebuilds can't mis-derive item_key and duplicate a post. - Rerun_Job_Support: queue/lock/budget/status-transient plumbing; jobs ride the M1 status transient so the board shows queued->pickup->result. Agent-Session: engineer-prab-m3-20260612
10 KiB
10 KiB