192. Replanning is bounded, idempotent and never weakens its intent
Context
Re-evaluating an intent must not file a new issue each time, loop forever, or quietly lower the bar (2026-10-08, 10-11, 10-16).
Decision
A drafted work item's identity is derived from (intent, criterion, gap kind); a replan reuses done work and supersedes through existing transitions; a policy-declared budget of drafts without a newly satisfied criterion ends in human_needed; a draft narrowing a criterion is refused without a recorded decision.
Alternatives rejected
- Work until satisfied
- An LLM noticing the same gap and filing a new issue
- Rewriting history on replan
Consequences
One budget field in policy.yaml; no runtime or scheduler; /proceed (ADR 0177 (on an unlanded branch)) inherits the bound.
Sources
Promoted on 2026-10-18 from the owner's ChatGPT project (workspace chatgpt-majordomus, issue I3100). The owner has not accepted it; it is proposed for the owner's decision.
chatgpt-majordomus:c196#3chatgpt-majordomus:c196#6chatgpt-majordomus:c205#2chatgpt-majordomus:c221#2chatgpt-majordomus:c270#8