Skip to main content
The Replay Hub lives at /replay (top-level sidebar entry). It’s the org-wide ledger of every local replay an SDK client has ever submitted via POST /api/v1/replay/{run_id}/results — useful for auditing replay coverage, spotting divergences across deployments, and tracking how much money replay has saved.

​
Why local replay?

Re-running a recorded run through the platform would cost real LLM dollars. The SDK’s replay_mode(ledger_path=...) instead executes the run locally, replaying recorded outputs from the run bundle. After the replay finishes, the SDK posts a summary — what was replayed, what diverged, how much was saved — to the platform. This page is where those summaries land.

​
Layout

The page has three sections:
  1. Header — title, total submission count, and a manual Refresh button.
  2. Stat tiles (4) — Submissions, Actions Replayed, Divergences, Cost Saved.
  3. Submissions table — most recent first, with columns:
ColumnDescription
Run IDOriginal run that was replayed (links to the run page)
Statussuccess / partial / failed — color-coded badge
Replayed / TotalActions replayed vs total in the bundle
DivergencesCount of comparisons where inputs or outputs differed
SavedUSD saved vs re-running through the platform
SubmittedWall-clock timestamp

​
Reading the stats

  • Cost Saved sums cost_savings across all submissions — that’s the dollar amount you’ve avoided by replaying locally instead of re-executing.
  • Divergences > 0 means the SDK observed a difference between recorded and replayed values. Common causes: timestamps in inputs, random seeds, UUIDs, time-dependent API responses. Click a row’s run ID to investigate at the per-action level on the run page.
  • Status partial means some actions replayed cleanly while others diverged.

​
Underlying API

  • GET /api/v1/replay/history/?skip=&limit= — paginated org-wide history (newest first; max 200 per page)
  • GET /api/v1/replay/history/{run_id} — submissions for a specific run
  • POST /api/v1/replay/{run_id}/results — what the SDK calls after ReplayMode.__exit__

​
When to use this page

  • Audit replay adoption — how many of your runs have been replayed locally? Are CI gates running replays as expected?
  • Catch determinism regressions — a sudden uptick in divergences after a deploy is a signal your run inputs/outputs grew non-deterministic dependencies.
  • Quantify savings — the Cost Saved tile is the headline number for replay’s ROI.

​
See also