- The same cached schema must localize pl and en independently, including Accept-Language parent fallback and raw keys
- YAML option maps keep order and scalar type, and a method provider is required at boot
- Strict config_form and fields documents keep source order and JSON scalar types
- Unknown keys, partials, path escape, and a mismatched modelClass fail activation with plugin context
- List-only controllers still activate when config_form.yaml is absent
- All locked field kinds, empty and single documents, and source order fail closed
- Unknown keys, types, duplicates, path escape, modelClass, partials, and missing assets must name the plugin, controller, and file
- Refresh, logout, and me use a separate PostgreSQL jti blacklist and safe profile
- Login stamps last_login only after a successful check and throttles repeated attempts
- Add the admin jti table, reset cutoff, and Winter indexes without AutoMigrate
- Reapply the developer and publisher seed idempotently and allow repeated role codes
- Fresh migrate is missing tokens_valid_after and the admin blacklist table
- Reapplying the seed is not idempotent and role codes reject Winter duplicates
- Same-secret backend token is still accepted by the frontend guard
- Permission denial must not invoke the schema or database callback
- Admin error bodies must not echo secrets or raw tokens
- Audience-aware mint, verify, refresh, and backend guard keep frontend tokens compatible
- Cabana mounts raw admin login, list schema, and record list behind admin.jwt.secret
- Framework migration seeds Winter backend users and developer/publisher roles
Documents the 103-method audit closure, the self-performed 11/11-closed
security review (with disclosure), the real defects check-phase8.sh's first
end-to-end run found and fixed, and the checkpoint decision to close Phase 8
with the Playwright UI matrix gap carried forward.
scripts/check-phase8.sh's final gate ran once with every stage green except
stage_ui_harness's Playwright browser matrix, a deliberate fatal() never
authored by 08-05. The user approved closing Phase 8 with this gap carried
forward; 08-VALIDATION.md flips 08-W0-07 green, marks 08-W0-08 partially
verified, and sets nyquist_compliant: false honestly. deferred-items.md
records what the follow-up spec needs to do.
Two real defects surfaced by the gate's first live run against the real
fonoteka-mcp SDK:
- stage_revoke looked up the connected app by a.name; ConnectedAppsIndex
actually serializes client_name (confirmed against
controllers/api/connected_app_controller.go serializeConnectedApp).
- Even with that fixed, stage_revoke ran after stage_replay, by which
point RevokeLineage's forward walk (presenting the pre-refresh spent
secret) had already cascade-revoked the live post-refresh access token
too -- correct, intentional T-08-REFRESH-REPLAY behavior, and the exact
same effect 08-09-PLAN.md's own mcp-lifecycle fixture ordering already
documented ('connected-apps would already be empty if list ran after
replay'). stage_refresh now captures the connected-app id while the
session is still live; stage_revoke DELETEs that id directly instead of
re-listing (ConnectedAppsDestroy has no revoked_at filter on its own
lookup, so this still exercises the real endpoint, idempotently, against
the id the real MCP-driven session actually owned).
08-10 Task 3 is the first time this gate has actually been executed
against real Docker/Postgres/the real fonoteka CLI/the real fonoteka-mcp
process. Four independent, previously-undetected defects surfaced:
- stage_postgres never set POSTGRES_INITDB_ARGS for the ICU pl-PL locale
lagoon.Use requires (every other Postgres testcontainer in this project
already does); the app failed to boot at all.
- stage_app_boot's seed step POSTed to
/_fonoteka/api/v1/onboarding/bootstrap, a route routes.go never mounts
(its own comment marks that group deliberately empty, pending a later
phase). The gate's test user/collection are now seeded directly with
SQL, matching every app-level OAuth test's own real-Postgres seeding.
- phase8_workdir() assigned PHASE8_WORKDIR from inside a function body
that is always invoked via command substitution (a subshell): the
assignment never escaped back to the calling shell, so every separate
caller (stage_postgres, stage_app_boot, each phase8_mcp_stage call, ...)
minted its own fresh mktemp directory. This silently fragmented one
run's state (app.log, the MCP client's gate-state.json) across dozens
of directories that never saw each other's writes -- the MCP client's
dcr stage could never see discovery's saved metadata. PHASE8_WORKDIR is
now set once, directly, in run_full_gate before any stage runs.
- gate-state.json (the MCP client's shared cross-invocation state) holds
raw live secrets by design and is never redacted; stage_secret_scan
correctly flagged it. It is now deleted once the MCP lifecycle stages
are done with it, before the scan runs -- the scan itself stays exactly
as strict as it already was.
stage_security_review also now refuses a nonzero threats_open count or a
missing required T-08-* row, not just a missing/unverified file, and gains
--security-review-only, a focused mode for Task 2's own verify command.
08-SECURITY-REVIEW.md: status: verified, 11/11 T-08 threats closed,
0 open, 0 accepted risks. Performed directly by the 08-10 executor
(no Task/Agent tool available this run, per the plan's documented
fallback) with re-executed named-test evidence for every threat; found
and fixed one real gap during the review (see the paired fix commit).
08-VALIDATION.md: 08-W0-01 through 08-W0-06 flip to green with their
automated commands re-run; nyquist_compliant and wave_0_complete are
now true. 08-W0-07/08-W0-08 (the full scripts/check-phase8.sh gate)
stay pending until 08-10 Task 3 actually executes it.
stage_security_review now also refuses a nonzero threats_open count and
any missing required T-08-* threat row, not just a missing/unverified
file. Adds --security-review-only, a focused mode running just this
stage (Task 2's own verify command) with no services booted.
Auditable one-to-one map of every PHP OAuth functional/security test
method to its named Go test/subtest evidence, mechanically verified by
fonoteka.go/parity/oauth_audit_test.go's TestPHPTestMap.
- register loopback http:// acceptance, javascript: URI rejection, and
error-body no-secret/no-stack-trace evidence
- token-exchange plain PKCE rejection even when verifier equals the
stored (non-S256) challenge
Fills in every scripts/check-phase8.sh stage skeleton with real logic:
disposable Postgres (docker run + pg_isready), the assembled Go app built
and served against it with a throwaway onboarding-seeded gate account,
the real unchanged fonoteka-mcp process started with all three required
environment variables, and the full scripted SDK lifecycle -- discovery
(MCP's own RFC 9728 401 hint, verified separately from authorization
server metadata), DCR, PKCE authorize, JWT login/consent, token, an MCP
tool call, refresh, replay of the spent refresh token, revoke, and a
post-revoke refresh failure -- delegated to the new
scripts/check-phase8-mcp-client.mjs driver, which resolves the MCP SDK's
auth helpers from fonoteka-mcp's own node_modules (no new dependency,
same pattern as parity/capture_clients.mjs). Both repositories'
vet/test/race, the full parity/corpus/secret-scan gate, the existing
check-phase8-ui.mjs --final-gate UI harness, an unchanged-client git-diff
check for both MCP_ROOT and NUXT_ROOT, and a 08-SECURITY-REVIEW.md
status:verified gate close out the stage list.
--contract-self-test validates structure only (stage names/order,
cleanup trap, loopback-only binding, the three MCP env vars, the
redaction helper, no pre-final full-run flag, read-only unchanged-client
references) in well under 30 seconds -- it boots no services. The
--red-contract self-test from Task 1 is preserved unchanged. run_full_gate
(the no-flag invocation) is 08-10 Task 3's sole execution site; 08-09
never invokes it.
Recording the full mcp-lifecycle fixture against real isolated PHP
(08-09-PLAN.md Task 2) uncovered three byte-level gaps between wristband's
assumed contract and actual production PHP behavior:
- Every explicit "Cache-Control: no-store" PHP sets is actually delivered
as "no-store, private" (Laravel's session-cookie default merges "private"
onto any explicit value); wristband's own default for unheadered JSON
error responses is "no-cache, private" (matching the house convention
already used elsewhere), not empty.
- PHP's redirect responses (authorize success and every error redirect)
render Symfony's default HTML redirect body with Content-Type
"text/html; charset=utf-8"; Go's bare 302 with no body never matched.
wristband/redirect_html.go ports that exact byte template, including
PHP's htmlspecialchars(ENT_QUOTES) escaping (Go's html.EscapeString uses
different quote entities).
tide/normalize.go: isIDKey now also masks "_ids" plural array fields
(e.g. collection_ids), a latent parity-corpus gap no prior fixture had
exercised with a literal, non-empty, non-placeholder array value.
- Declares the ordered Phase 8 stage list and stage function skeletons
- --red-contract <stage> is a permanent RED-harness self-test hook
(exit 86, PHASE8_STAGE:<stage>:FAIL:PHASE8_RED:real-mcp-stage)
- --contract-self-test and the full gate are completed in Task 3/08-10
- bonfire.wrap registers a Repeatable Flag as a Cobra StringSlice so
Input.Flags returns every repeated --name=value occurrence in order;
scalar/bare flags are unaffected (D-19)
- wristband.IssueClientCredentials/RejectRedirectURI export the exact
random-id/secret/hash and redirect-URI validation RFC 7591
registration already uses, so the fonoteka:oauth-client operator
command shares one hash/validation path with DCR (T-08-SECRET-TIMING)
- TestPhase8RedBonfireFlags asserts Input.Flags preserves ordered
repeated --redirect-uri/--scope values while existing scalar --mode
flags via Input.Flag stay unaffected
- Adds the compiling seam (Flag.Repeatable, Input.Flags,
cobraInput.Flags) without wiring Cobra StringSlice registration yet,
so the test fails with PHASE8_RED:bonfire-flags (D-19)
rotateRefreshToken ports OAuthCodeManager::rotateRefresh: a fresh refresh
token rotates atomically (revoke old access token, mint successor, link
rotated_to_id) while a replayed (already-rotated) token instead revokes the
whole lineage and commits that kill before Token maps it to invalid_grant
outside the transaction (T-08-REFRESH-REPLAY). Token also runs the D-17
expiry sweep (DeleteExpiredCodes/DeleteExpiredRefreshTokens) before grant
processing. Server.Revoke is the new cascade-revoke seam a connected-app
controller uses instead of touching refresh rows directly.
TestPhase8RedLifecycleFramework drives exchange -> rotate -> replay against
the real (in-memory-backed) Server.Token and fails while rotateRefreshToken
is 08-04's invalid_grant placeholder (PHASE8_RED:lifecycle-framework,
verified fail-closed via scripts/check-phase8-red.sh). Extends the
RefreshTokenStore/AuthCodeStore interfaces with the store seams Task 2's
implementation needs (ByAPITokenIDForUpdate, MarkRotated,
DeleteExpiredCodes, DeleteExpiredRefreshTokens) and updates the framework's
in-memory test double to satisfy them.
TestFetchTooLargeIsStreaming asserted the server wrote at most 64 KiB, but
the handler keeps flushing 64-byte chunks until the client's close propagates,
which under a loaded full-suite run exceeds that (observed ~80 KiB). The
assertion guards against unbounded buffering toward 8 MiB, so 1 MiB keeps
the intent and removes the flake.
check-phase8-ui.mjs encodes 08-UI-SPEC.md's full consent/connected-app
state matrix, accessibility, responsive, and i18n contract as a versioned
32-scenario catalog across 7 categories. --contract-self-test validates
catalog completeness, guarded Nuxt source-file hashes (proving the
harness itself never writes inside vue-fonoteka-app), and that
@playwright/test resolves from the already-installed dependency, all
without booting a browser or service (runs in ~50ms).
--final-gate (running verify:oauth-return-path, verify:oauth-i18n, and
the real Playwright matrix) is scaffolded but refuses to run without
PHASE8_UI_ALLOW_FINAL_GATE=1 and is explicitly 08-10's closing-checkpoint
responsibility, not executed by this plan.
Server.PendingRequest/IssueCode/DenyPending port PHP
OAuthConsentController::pendingFor/OAuthCodeManager::issueCode as
app-agnostic protocol operations (08-CONTEXT.md D-08): every missing,
foreign-owner, used, expired, or already-issued pending row collapses to
the identical ErrPendingNotFound (T-08-CROSS-USER/T-08-REQUEST-LEAK).
IssueCode trusts the caller's already-computed granted scopes/collection
ids and returns the ordered redirect_to URL built through the existing
RFC 3986 encoder.
AuthCodeStore.MarkIssued gains scopes/collectionIDs/expiresAt parameters
(PHP's issueCode overwrites all three, not just code_hash/user_id) and
ClientStore gains MarkConsented, both required for D-08's consented_at
stamping and server-derived grant persistence. Options gains CodeTTL
(600s PHP-parity default) following the established Options-extension
pattern.
- Server.Token: JSON rejection before ParseForm, body-over-query precedence,
Basic-over-form client auth, exact invalid_request/unsupported_grant_type/
invalid_client/invalid_grant bodies, Cache-Control/Pragma on success only
- authenticateClient: public/confidential dispatch, constant-time secret
compare (T-08-SECRET-TIMING)
- exchangeAuthorizationCode: single WithinTx lock/consume/mint/refresh-create
covering code/client/redirect/resource/PKCE binding and single-use replay
(T-08-CODE-REPLAY), sequential and concurrent proofs
- rotateRefreshToken: grant_type=refresh_token dispatches per PHP validity
but is a deliberate invalid_grant placeholder; full rotation is 08-06
- full token_test.go behavior matrix appended alongside the RED anchor
