Skills need a frontmatter opt-out from post-compaction replay (stale $ARGUMENTS re-executed; real-world unintended git push)

Summary

Post-compaction skill replay hands the model incomplete, context-stripped instructions — and skills have no way to opt out of it. After compaction, Claude Code re-attaches every previously invoked skill as an invoked_skills context block — the rendered body, with the user's $ARGUMENTS already substituted in, truncated to the first 5,000 tokens per skill (25,000 combined; documented at https://code.claude.com/docs/en/skills). The replayed arguments arrive stripped of the conversation that gave them meaning, so the model can verify neither their freshness nor their referents — which turns a replayed line into unpredictable, undesired behaviour. No frontmatter field or setting excludes a skill from this; disable-model-invocation: true has no effect on it.

That combination is the core problem: some skills are one-shot by design — e.g. an /ask skill ("answer this question directly, don't implement") whose invocation is meaningless once its turn ends — and the author knows that, but today cannot express it.

Real incident

Hours-old /ask arguments replayed after compaction contained a push request. The model executed it as a live instruction and resolved it against the wrong repository — the request had meant repo B, but the compacted context had stripped the referents that made that clear, so repo A was pushed. The block's own "NOT the user's current message" disclaimer did not prevent re-execution.

This is the same failure mode as #50724 ("stale skill ARGUMENTS re-surfaced as fresh prompt") and #61848 (its duplicate), both closed NOT_PLANNED — respectfully, that seems like the wrong resolution for a defect class that produces unintended side-effectful actions (pushes, commits, message sends) from stale input. Adjacent open issues (#83306, #72549, #82144, #82851) concern the same attachment mechanism but none covers the missing opt-out / argument re-execution case.

Requested fix

A frontmatter opt-out:

compaction-replay: false

so one-shot skills that embed raw user text can be excluded from the replay entirely — the fix that matches the semantics, decided by the skill author. An arguments: ephemeral variant (replay the instructions, never the substituted arguments) would serve skills whose body is worth keeping. Secondary mitigations: omit substituted arguments from the re-attached copy (suggested in #50724), or timestamp-tag them so staleness is prominent.

Environment

Claude Code 2.1.226, Linux; long session, user-invoked one-shot skill with $ARGUMENTS interpolation.