[justinmclean]

Summary

• The spec-loop plan beat derived work items from Known-gaps bullets without checking they were still true, so a bullet describing shipped code sent the loop at finished work (the drift specs/spec-gap-staleness.md documents). This is the prompt half of that spec, acceptance criteria 1 and 2.
• PROMPT_plan.md: a gap becomes a work item only with evidence it is still open: a call site, file content, or a counted value, written into the item as a *Verified open:* line. A name match is explicitly not evidence. An unverifiable bullet is neither queued nor deleted but listed under Notes & discoveries as unverified. A bullet with a gap-closes-when predicate defers to spec-validate.
• PROMPT_update.md: every Known-gaps bullet the update beat keeps must be re-verified the same way; a closed one is removed or rewritten, an unverifiable one is kept and marked, and a gap with a crisp end state gets a predicate.

Plan item plan-beat-evidence-rule.

Type of change

• CI / dev loop (prek, workflows, validators)
• Other: spec-loop prompts

Test plan

• The plan item's validation: grep -qi evidence on both prompts, bash -n tools/spec-loop/loop.sh.
• New fixture test test_prompts_carry_the_evidence_rule pins the three evidence kinds, the predicate syntax, and the no-name-match rule in both prompts. Mutation-checked: with the old prompts it fails.
• bash tools/spec-loop/tests/test_runner_fixtures.sh passes; shellcheck findings unchanged from main.
• prek run --all-files --skip workspace-pytest: every hook passes except skill-token-count (tokenizer download blocked in the authoring sandbox; no SKILL.md touched). RAT: 0 unapproved.

RFC-AI-0004 compliance

Not applicable (prompt text only).

Linked issues

Companion to the spec-validator-gap-checks PR, which builds the gap-closes-when checks these prompts refer to. Independent of it: neither touches the other's files.

Notes for reviewers (optional)

• spec-gap-staleness.md is deliberately not edited here; the validator PR rewrites its Known-gaps bullet, and editing it in both would conflict. Once both land, the spec can move from proposed to experimental.
• The reconciled plan in the open spec-sync PR already uses *Verified open:* lines in this shape.
• Generative-AI disclosure: authored with Claude Code under maintainer direction; the commit carries the Generated-by: trailer.

[potiuk]

Thanks — the evidence rule reads well and matches the predicate vocabulary #1940 shipped. I rebased the branch onto main and pushed one small fixup commit, 86052d8 (please pull before pushing again):

• tools/spec-loop/specs/spec-gap-staleness.md — with #1940 merged and this PR putting the rule in both prompts, the "evidence rule in the prompts is not built" Known gap described shipped work, the exact staleness the spec exists to prevent. Removed, and the spec moves from proposed to experimental (here and in the overview table), as the PR body anticipated once both halves landed.
• PROMPT_update.md — symbol-called reads Python sources only (a directory without *.py is invalid), so the update beat now uses path-exists or none (narrative) for a gap about a shell function, which most of spec-loop is.

GitHub has not yet computed whether this branch merges cleanly with its base.

---

This review was drafted by an AI-assisted tool and
confirmed by an Apache Magpie maintainer. The maintainer
approving this PR has read the findings and signed off. If
something feels off, please reply on the PR and a maintainer
will follow up.

More on how Apache Magpie handles maintainer review:
Contributing guide.