The probe workbench (/bench)
ply-web runs ply's probe scenarios (ply/probes, pinned in package.json) against Cloudflare Artifacts from the Worker, and shows each run. ply is the client under test; this Worker is the probe runner; Artifacts is the server under test. All of /bench is private: send Authorization: Bearer <PLY_TOKEN>, or sign in through the browser's Basic prompt with PLY_TOKEN as the password (any user name).
GET /bench: the catalogue (each scenario: runnable here, stock git only, or a seed too large for a Worker), the run form, recent runs.POST /bench/runs(same-origin form:scenario,seed): one run. At most one at a time and 20 an hour. Answers with a redirect to the run once it is over.GET /bench/runs/<id>: the verdict, the steps, the end state, the protocol waterfall, every exchange (status, timings, bytes, cf-ray, headers, pkt-line commands, advertised refs, report-status, object ids, the fault applied), the cleanup counts and the run's spans.GET /bench/runs/<id>/exchanges/<i>/request|response: a raw body (R2, at most 1 MiB).GET /bench/runs/<id>/<slug>.jsonand<slug>.test.js: the replay for ply's suite. Put both files intest/probes/replay/of a ply checkout and runnpx vitest run test/probes/replay/<slug>.test.js. That runs the plan with ply and stock git against git http-backend (the reference), and checks the recorded run with that commit's evaluator.
A run
- Build the seed in memory, with the same object ids as ply's runner (
src/bench/seed.ts). - Create
pb-<seed>-<run>in theply-loadgennamespace (bindingPROBE_REPOS) and seed it. - Run each step with ply through the recording fetch (
src/bench/recorder.ts), applying the step's fault. Read the result back with ply (src/bench/judge.ts); there is no fsck and no merge bases, because a Worker has no git. - Classify (
src/bench/verdict.ts) each difference from the stated expectation:client_bug: ply's outcome contradicts the report-status of the same exchange.server_bug: the server confirmed an update its refs do not hold.expected_incompatibility: an entry in ply's allow-list.unresolved: anything else (export it to close the triangle).
- Clean up. Revoke every token of the repository, including the one create minted, and prove the revoked token gets 401 or 403 (cf-ray recorded). Then delete the repository and prove it is gone. ProbeBench's alarm sweeps any repository a run did not prove cleaned.
State
- ProbeBench (SQLite Durable Object, via
ctx.exports) is the source of truth: runs, the repositories it created, and admissions. - R2
ply-benchholds the exchange bodies.
Nothing is cached: every page is private and no-store.