Replay Testing

replaytest captures live API traffic into a scenario file, then replays it by serving the recorded responses from an in-page web worker. Because the worker answers the app's own fetch and XHR calls in-process, front-end runs are fully local and deterministic — no real backend, no network flake, no mock server to keep in sync by hand.

On top of plain replay, fuzz mutators perturb the recorded responses so you can see how the UI handles malformed or degenerate data, and explore splits a scenario into seeds so the calling agent can fan out subagents for breadth.

Pro feature

Record, refine, replay, and explore require a Pro license with the advanced_testing capability (agnt activate <key>). list and show are free. Recording is currently alpha. See Licensing below.

The loop

start proxy → record → drive the app → stop → refine → replay across presets → explore

replaytest {action: "record", name: "checkout", proxy_id: "app"}
// …click through the flow in your browser…
replaytest {action: "stop", name: "checkout"}
replaytest {action: "replay", name: "checkout"}
replaytest {action: "replay", name: "checkout", fuzz: "empty_array"}

Actions

Action	License	Description
list	free	List scenarios in .agnt/replaytests/
show	free	Show a scenario or its last run report
record	Pro	Capture API traffic + interactions against a running proxy (requires daemon mode)
stop	Pro	Finalize the recording into .agnt/replaytests/<name>.json
refine	Pro	LLM-assisted scenario cleanup (needs ANTHROPIC_API_KEY or CLAUDE_KEY)
replay	Pro	Drive headless Chrome, mock the network in-page, assert via DOM signature
explore	Pro	Return a seed partition for fanning out browser-debugger subagents

Scenario files

Everything lives under the project, so scenarios are reviewable and committable:

• .agnt/replaytests/<name>.json — the scenario (recorded requests/responses plus assertions)
• .agnt/replaytests/<name>.report.json — the most recent run report

Fuzz presets

replay can mutate the recorded responses before serving them, which is the cheapest way to find the error paths nobody wrote a fixture for:

Preset	Effect
empty_array	Replace array payloads with []
http_error	Return an HTTP error status instead of the recorded 2xx
truncated_json	Cut the JSON body short
null_fields	Null out object fields
reordered	Reorder array elements and object keys
type_flip	Flip value types (string ↔ number, etc.)

Licensing

agnt Pro uses self-hosted, offline license validation — the binary embeds only a public key and verifies the signature locally. There is no runtime phone-home.

agnt activate <key>        # validate + install a license blob, offline
agnt license status        # state, email, expiry, days left, capabilities
agnt license deactivate    # remove the installed license

The dividing line is breadth of operation, not a fixed tool list: single-page operations (responsive/API/loading/accessibility audits, snapshots, browser debugging, process and proxy management) are free; whole-site, multi-page, and application-wide work is Pro. A license that has passed its expiry keeps working for a 14-day grace period, with a renewal warning, before Pro capabilities stop.

The license blob is stored per-user at ~/.local/state/agnt/license.lk ($XDG_STATE_HOME/agnt/license.lk), written atomically at mode 600.

Current limitations

• Recording is alpha and requires daemon mode; the non-daemon path returns a clear "requires daemon mode" message rather than half-recording.
• refine needs ANTHROPIC_API_KEY or CLAUDE_KEY at runtime.
• Orphaned blobs from coalesced duplicate recordings are retained in the scenario JSON. Harmless, but a future GC pass will prune them.