whenpoem Replay
Replay is an inspection over frozen state. It does not execute new experiments and must not modify live graph nodes, ratings, judgments, or preregistrations.
Workflow
- If the user did not provide a snapshot id, call
mcp__memory__list_snapshots(limit=20). Show label, timestamp, and counts; do not create a new snapshot merely to replay the past. - Phrase exactly one counterfactual in a falsifiable sentence, such as
Approve hyp_07 instead of pausing it. - Call
mcp__memory__replay_counterfactual(snapshot_id, counterfactual)and retain the returnedbranch_id. - Inspect only the recorded snapshot and branch metadata. Do not call model
scripts,
seed_perturb, or the sequestered-data query tool. - If the analysis suggests a live action, report it as a recommendation. The
user may later call
resume_branch; replay itself never flips status.
Output shape
Return:
snapshot_id: <snap_...>
branch_id: <replay_...>
counterfactual: <one sentence>
live_graph_mutated: false
evidence: <recorded facts used>
finding: <what changes under the counterfactual>
recommendation: <none or explicit user action>
limitations: <what replay cannot establish without a new experiment>
For several what-ifs, create one replay branch per counterfactual and keep each
output independent. Use mcp__memory__list_replay_branches(limit=20) to audit
existing branches before adding duplicates.
Completion criteria
Replay is complete when the source snapshot and new branch are identifiable, the live graph is unchanged, every finding cites recorded snapshot evidence, and any proposed live change is clearly left for separate user approval.