Testing: how to verify a Poppy implementation

Testing: how to verify a Poppy implementation

Testing Poppy is mostly a discipline of layers: a static schema pass proves very little, the official validator has real side effects, and community claims prove nothing until you run them. This page orders the ladder and records exactly where this site’s coverage stops.

Layer 1 — Static schema checks

Feed the discovery document through the official JSON Schema / Validator (kit/company in the official repository). This catches structural errors — missing required fields, wrong types, broken references.

What it proves: the document is well-formed against 1.0. What it does not: that discovery, sessions, OAuth, APIs or revocation work at runtime. A schema-valid document with a dead issuer is still schema-valid.

Layer 2 — The official validator, with its side effects understood

The official validator ships two modes with different blast radii:

  • Survey (GET-only) — read-only checks against a deployed endpoint. Safe to run against any domain.
  • Full check — acts as a personal agent: it creates sessions and exercises flows. Not side-effect free. Run it only against endpoints you are authorized to use.

This site’s recorded run: the validator’s own suite plus a read-only survey against a local, schema-valid fixture (fixtures/company-survey/, 121/121, 2026-11-17 baseline), which proves static discovery conformance for the fixture and nothing beyond it.

sh fixtures/company-survey/run.sh   # official validator suite + survey, offline fixture

Layer 3 — Session, OAuth and revocation — partially tested

Direct Sign-In now has a request-level rehearsal on record (fixtures/denied-flow/, 2026-11-17): a local HTTPS Company fixture implements the OAuth endpoints (PAR → authorize → token → API) and a fixture Personal Agent drives both outcomes of the consent page.

  • Denied branch: the redirect back carries error=access_denied with the issuer in iss (RFC 9207, required even on denial by spec 4.2) and the round-tripped state — and the audit counter proves it minted no code and no token;
  • Approved branch: PAR with a PS384 client assertion (4.1) and a DPoP proof, PKCE S256, code exchange, bearer token, protected API read; a forged code gets invalid_grant and an unknown bearer gets 401.
sh fixtures/denied-flow/run.sh   # deny + approve + negatives, offline fixture

What is simulated: the consent-page choice arrives as a fixture parameter instead of a browser, and DPoP proofs are required present but not signature-verified. Sessions (per-request Poppy-Session JWT), token refresh and revocation (the three distinct semantics) still have no run on record. When you test them:

  1. Test sessions before login — a signed-out session is a legal state;
  2. Test login separately — approved scopes must match token contents;
  3. Test revocation last — refresh-token revocation, session deactivation and account sign-out are three different operations; verify each does not silently do the others'.

Reading community claims

The capability matrix records what community implementations claim at their pinned revisions — with the evidence location. A claimed cell is the author’s statement, not your result. Before relying on one, re-run the relevant layer yourself at the revision you will actually deploy; none of the community rows targets the official 1.0 as of this site’s review, so treat their 1.0 compatibility as unknown.

What this site has run — and has not

LayerStatus
Official validator suitetested (121/121, offline fixture)
Read-only survey vs. fixturetested
Direct Sign-In incl. user denialtested (request-level offline fixture, 2026-11-17)
Sessions / token refresh / revocation / DPoP signature validationnot tested — no live endpoint
Any community implementationnot run

Keep the three evidence states separate in your notes: what you ran, what the authors claim, and what nobody has evidence for.

Last updated on