Runs and replay

Every time your agent runs — local dev or Agent21 production — CodeNow captures the entire interaction as a run: a sequence of events stored as JSONL in .codenow/runs/{slug}/{ts}-{runId}.jsonl.

What’s in a run


{ "kind": "run.start", "runId", "slug", "ts", "input", "kit", "model" }
{ "kind": "message", "ts", "message": <raw SDK message> }     ← assistant text
{ "kind": "message", "ts", "message": <raw SDK message> }     ← tool_use
{ "kind": "message", "ts", "message": <raw SDK message> }     ← tool_result
…
{ "kind": "run.end", "ts", "durationMs", "status", "resultMessage" }

The recorder shipped in every CodeNow-scaffolded agent appends these as the run executes. JSONL is append-only — never has to rewrite an existing file.

Storage caps

To keep .codenow/runs/ from growing forever, the recorder prunes on every new run start:

100 runs per slug
500 runs total across all agents in the project

Pruning is async and best-effort — it never blocks an agent run.

Replay viewer

Open the Runs sub-tab in the Agents panel. Every recorded run is listed newest-first with cost, duration, status, source (local / production). Click any row to open the Replay viewer:

Step-by-step timeline — colored icons per event kind (input · system init · assistant text · tool call · tool result · result · error · end)
Keyboard nav — j/k to step, Space to expand the current step’s details, Esc to close
Tool call args + tool results expand inline as collapsed <details> blocks

Fork to current code

The screenshot Cursor can’t ship: take any past run, hit Fork to current code →, and CodeNow re-runs the same input through your current local agent code (spawning tsx agent.ts "<input>" or python3 agent.py "<input>"). Then it shows the original and the fork side-by-side:

✓ “Output matches” if the assistant text is identical
⚠ “Output differs” otherwise — with cost/duration/step-count deltas

This is how you verify “my refactor didn’t change behavior” — replay yesterday’s prod incident through today’s code.

Production runs

When you connect Agent21, the Runs tab also shows production runs from the marketplace. They merge into the same chronological list with a prod badge. Production runs are read-only — you can replay them but not Fork (since they ran on Agent21’s infrastructure, not yours).

See /concepts/manifests for how the production manifest preserves provenance.