Replace the Python/kaya server in server/ with a Rust Cargo workspace of
three crates mirroring the old layout: tavolo-platform (lobby API, live-game
store, websocket endpoint, deadline scheduler, stats/Elo, OIDC sessions),
tavolo-scopone (pure scopone scientifico rules plus the platform adapter)
and tavolo-app (composition root and the `tavolo` binary with
serve/migrate/backfill-elo).
The HTTP and WebSocket contracts, error messages, Redis key layout and
Postgres schema are unchanged, so the existing web/ frontend and in-flight
games keep working. Notable differences: sessions are JSON under
tavolo:session: (Python pickle sessions invalidate once), LOGGING_CONFIG is
ignored in favour of RUST_LOG, DATABASE_URL must be Postgres, sqlx migrations
replace aerich, and OpenTelemetry exports Rust process metrics instead of
CPython gc metrics (no Postgres/Redis client instrumentation).
Docker Compose and the Kubernetes manifest now run the single Rust image
(embedded migrations, /static and /assets served by the app); the release CI
keeps building server/Dockerfile. The full test suite (scopone rules,
platform store/deadlines/stats/websockets, OIDC login flow, scopone e2e) is
ported and runs with cargo test --workspace; Redis/Postgres integration tests
are gated on TEST_REDIS_URL/TEST_DATABASE_URL.
kaya-oidc 0.0.6 can run several providers in one app (OIDCProviders),
each with its own login route and namespaced session state. Adopt it:
providers are configured in indexed, contiguous OIDC_<n>_* blocks
(0-based) with OIDC_<n>_PROVIDER naming the provider in its login URL
and session keys. The old unindexed OIDC_ISSUER/OIDC_CLIENT_ID/... names
are no longer read; startup fails with a rename hint when they are set.
With no OIDC variables at all, the single dev mock provider named
"default" is used, as before.
The platform now authenticates through OIDCProviders: auth.current_user
returns the session's user (sessions stay single-identity, so at most one
provider is logged in) and require_auth/route handlers use it, while the
websocket path keeps scanning for any provider's user. New routes expose
the configuration to the SPA and keep old links working: GET
/api/auth/providers lists the login URLs, GET /auth/login redirects to
the first provider and GET /auth/logout to whichever provider the
session is logged in with. The lobby renders one login button per
provider from that endpoint.
Update the env examples, compose and k8s manifests to the indexed scheme,
document the multiple-provider setup and the migration, and cover it with
config, auth and route tests.
kaya-oidc 0.0.6 namespaces its session state per provider as
oidc:{provider}:user (the default provider is "default", so the key is
oidc:default:user). The unprefixed oidc_user key is no longer written:
left unadapted, every websocket upgrade would be rejected with 401 and
traced requests would silently lose their user.id attribute. Adopt it
with a session_user helper on tavolo.platform.auth that finds any
provider's user, used by both get_ws_user and telemetry's
session_tag_hook. Websocket and route handlers keep their signatures.
kaya-otel 0.0.6 only enriches websocket spans with standard URL, network
and peer attributes (plus error.type), which needs no application change.
Regenerate the lock: the eight kaya pins move to 0.0.6 and, the wheels'
declared metadata being unchanged, nothing else moves. tavolo-platform
now floors kaya-core>=0.0.6 (it reads the new session layout) and
tavolo-app floors kaya-otel>=0.0.6.
As kaya documents, the session-key move signs existing logged-in users
out once; they simply log in again.
kaya 0.0.5 surfaces a vanished websocket peer as
kaya.core.WebSocketDisconnect instead of leaking granian's
RSGIProtocolClosed: receive() reports a close message so iteration ends
normally, sends raise the new exception, and KayaApp.handle_websocket
swallows it as normal termination. Adopt it: drop the granian-specific
workaround from the platform websocket handler, catch
WebSocketDisconnect in the send helper, and model the new semantics in
the abrupt-disconnect test (with coverage that a send on a vanished
peer is ignored).
Regenerate the lock: the eight kaya pins move to 0.0.5 (declared
metadata is unchanged) and pip-compile also bumps PyJWT to 2.15.1, a
security release with no breaking changes for kaya-oidc's jwt.decode
usage. tavolo-platform now floors kaya-core>=0.0.5 (it imports
WebSocketDisconnect) and tavolo-app floors kaya-otel>=0.0.5.
A close() before accept() carries the HTTP status of the refused
upgrade, not a WebSocket close code: RSGI (granian) uses the value
verbatim as the response status, so 4401 panicked on the invalid status
(InvalidStatusCode) and — granian being built with panic = "abort" —
aborted the worker and took the whole app down. ASGI servers answer 403
regardless, and browsers report a refused handshake as an abnormal close
(1006), so the 4401/4403/4404 codes never reached clients anyway.
Reject 401/404/403 to mirror GET /api/games/{id}, and classify a failed
first connection in the game page through the REST API so players still
see "session expired" / "game no longer available" instead of a retry
loop. Add a regression test asserting every pre-accept rejection uses a
status in the HTTP range.
The root logger runs at DEBUG, so urllib3.connectionpool emitted a
record per pooled HTTP request. Keep its INFO+ output while dropping
the per-connection chatter.
The history table truncated finished_at, which the API serializes in
UTC, so the Finished column always showed UTC wall time. Parse the
timestamp with the browser's Date and render its local components,
keeping the existing YYYY-MM-DD HH:MM layout.
Cursors are JSON, so a datetime sort key reached the next request as a
string. Tortoise's DatetimeField.to_db_value does not parse strings, so
asyncpg rejected the string where a timestamptz parameter was expected;
sqlite tolerated it, hiding the bug in tests. Coerce decoded cursor
values through their model field's to_python_value before building the
keyset filter.
Builds the TracerProvider in tavolo.telemetry (kaya-otel's is private) and
injects it into kaya-otel, then instruments asyncpg and redis through
opentelemetry-instrumentation-asyncpg/redis. Redis command values are
redacted and asyncpg spans carry the parameterized SQL only.
Adds tavolo.active_users and tavolo.ongoing_games gauges, process/CPython
metrics via opentelemetry-instrumentation-system-metrics and OTLP log
export. The game store tracks ongoing matches (Redis sorted set, in-memory
set) and the session stores expose live counts for the gauges; a poller
refreshes them since metric callbacks cannot await the stores.
The tavolo-app otel extra was never installed: requirements.txt was
locked without it and the Dockerfile installs the local distributions
with --no-deps, so any container with OTEL_ENABLED set died at import
with "kaya-otel is not installed".
Add the extra's full closure (kaya-otel plus its OpenTelemetry/protobuf
dependencies) to the pinned lock and refresh the comments that claimed
the extra had to be added separately.
GET /api/games/preview returns the lobby behind a code plus the game's
name and engine-echoed options. The lobby debounces the lookup as a
complete code is typed and shows the game, seated players, open seats
and option badges rendered from the game's schema, disabling Join when
the lobby is full or the match has already started.
With the rule enabled, an ace played on an empty table goes straight to
the player's capture pile instead of staying on the table: it counts as
the last capture but never scores a scopa. An empty capture request now
matches only that self-capture option, so requesting nothing still can't
satisfy a real capture.
An ace played when no equal rank is on the table sweeps the whole
table; the sweep never scores a scopa, and an ace already on the table
still forces the mandatory equal-rank capture. The rule is off by
default, selectable per match via the schema-driven creation form, and
announced in the game view.
Bootstrap 5.3 ships .card-img{width:100%} for images inside cards, which
collided with the similarly named class emitted by the card helpers:
every card image stretched to its container width, so the table cards
wrapped one per flex line and pushed the player's hand below the
viewport. Emit 'playing-card' instead and rename the six selectors in
style.css; callers keep passing only the size modifiers.
The Dockerfile copies the frontend inputs explicitly, so the new
theme.css and vendor/ directory were missing: trunk failed to resolve
them and the image build stopped in the web-builder stage. Copy both
into the build context; the runtime image is unchanged.
The lobby was hand-rolled CSS and looked plain. Vendor Bootstrap 5.3
(MIT) under web/vendor so builds stay offline, load it in dark mode and
bridge the felt/gold palette onto Bootstrap's CSS variables in a new
theme.css; style.css now only owns the game table, cards and scoring
summary.
Rebuild the lobby, waiting room and shared chrome with Bootstrap
components: a shared navbar, creation form as cards with schema-driven
controls, roster as a list group with seat/team badges, a seats progress
bar and a copy-to-clipboard button on the join code. History,
leaderboard, connection/error states and the 404 page use cards, tables
and buttons too; the play table itself is unchanged.
The copy feedback timer is cancelled when the waiting room is rebuilt by
another player joining, fixing a disposed-signal panic that froze the
page mid-match.
The deadline consumer runs in a long-lived task outside any request, so
a timeout that ended the match raised 'No TortoiseContext is currently
active' in save_match_result before the state was saved: the game stayed
stuck on the last turn and the entry retried forever. It only surfaced
on the match-deciding turn; ordinary timeouts and human plays were fine.
Bind the Tortoise context before persisting a finished match (optional
context_binder wired to TortoiseMixin.ensure_context), and back off to
the heartbeat when a due entry fails instead of hot-looping on it.
Relocate the composition root (app, config, static, logging_config,
aerich_config) from server/src/tavolo/ to
packages/tavolo-app/src/tavolo/, with its own pyproject.toml and README
like the sibling distributions. Module paths and the granian entrypoint
are unchanged.
server/pyproject.toml remains as a tooling shim ([tool.aerich] next to
./migrations, [tool.mypy] for runs from server/); the Dockerfile installs
all three distributions from ./packages. Verified: all suites green,
mypy clean per package, aerich resolves the moved config identically,
wheel builds.
Send creation options as the nested options object the lobby expects,
drive the creation form from each game's options_schema, and render
history/leaderboard from the generic result payload (nullable teams,
float points). The lobby uses the payload's seats_open, and the
scopone-specific table, hand-end and game-over views are gated on its
game_type with an unsupported-game fallback for future games. The
websocket protocol is unchanged.
Move the game-independent machinery (lobby, live-game store, websocket,
deadline scheduler, match history, leaderboards) into a new
tavolo-platform distribution behind a GameEngine contract, the scopone
scientifico rules plus a platform adapter into tavolo-scopone, and keep
only the composition root in tavolo-app. The three distributions share
the tavolo namespace (PEP 420, kaya-style monorepo).
Match history becomes fully generic: Match carries the engine's result
JSON and MatchPlayer points/details instead of scopone-shaped team
columns (migration 3 backfills existing rows). Lobby creation takes an
opaque per-game options object and websocket actions dispatch to the
session's engine.
Tests: platform suite runs against a DummyEngine toy game, scopone
keeps the rules tests plus new adapter tests, server/tests covers the
wired stack end to end (194 tests, was 143).
pwo.async_test runs every test in a fresh event loop and TortoiseMixin
builds one TortoiseContext per loop. Dropping the previous context when
the loop changed orphaned its aiosqlite connections; their non-daemon
worker threads then blocked threading._shutdown forever, so the suite
printed OK but the interpreter never exited.
Add TortoiseMixin.aclose() and a tests.helpers.async_test wrapper that
closes the context in a finally on the test's own loop, and switch the
test modules over to it.
Adds an OTEL_ENABLED opt-in that assembles kaya-otel's OTelMixin into
the app, exporting HTTP/WebSocket traces and request metrics to an
OTLP/HTTP collector. Configured through OTEL_SERVICE_NAME,
OTEL_EXPORTER_OTLP_ENDPOINT, OTEL_EXPORTER_OTLP_HEADERS and
OTEL_EXCLUDED_PATHS (defaults to the /api/health probe endpoint).
kaya-otel is an optional 'otel' extra imported lazily, so default
installs and the test suite do not need the OpenTelemetry packages.
The kaya dependencies are bumped to 0.0.4, which kaya-otel requires.
Each player's rating starts at 1500 and updates transactionally with
every finished match: a team's rating is the mean of its two members
and the standard K=32 formula decides the zero-sum delta applied to
both members of a team. Ratings are per game type in a new
player_rating table; match_player records each match's elo_delta.
- GET /api/leaderboard exposes elo and sorts by it
- GET /api/me/matches includes per-player elo deltas
- new GET /api/me/ratings returns the caller's rating per game type
- frontend: Elo column on the leaderboard, per-match delta in the
history page, current rating in the lobby
- python -m tavolo.backfill_elo recomputes all ratings from the
recorded match history (one-off backfill for existing matches)
The socket had no recovery path: a mid-game drop left the table showing
stale state with no indication, and plays were silently swallowed by the
dead channel. Surface the server close code from ws::connect and add a
reconnect driver in the game page: transient losses retry with
exponential backoff (capped, then a manual Retry), while deliberate
closes (session expired, game gone) stop retrying. Reconnects are free
resyncs because the server pushes a full state snapshot on connect.
Also gate card clicks while disconnected, show a connection banner, and
drop the ticking interval and pending retries on unmount.
The mixin received the event loop from Kaya but ignored it, calling
asyncio.get_running_loop() instead. Under granian RSGI __rsgi_init__ runs
before the loop starts, so that raised RuntimeError and killed the
worker. Use the loop passed to setup(), falling back to the running loop
for the lazy calls from sync_deadline().
Turn auto-play and hand-end auto-continue were process-local asyncio
tasks armed only by client connects and state broadcasts: with no
sockets connected the next turn's timer was never armed, a hand-end
timer died with its worker, and neither survived a pod restart.
Deadlines are now driven by the absolute timestamps persisted on the
game state and enqueued in a shared Redis sorted set. Every worker runs
a consumer that fires due entries under the per-game lock after
revalidating them against the live state, so timeouts no longer depend
on any player being connected and survive the death of any worker.
Delivery is at-least-once: entries are removed only after processing,
and revalidation makes duplicate deliveries no-ops.
Queue entries carry the deadline as integer epoch milliseconds, which
also serves as the revalidation token, and the score derives from the
same value.
Assemble the Postgres DSN from DATABASE_ENGINE/HOST/PORT/NAME/USER/
PASSWORD/OPTIONS so only the password needs to live in a secret; the
rest can go in a ConfigMap. DATABASE_URL remains a full-DSN override
(used by the sqlite test suite). Credentials are percent-encoded, the
port and options are omitted when empty, and the k8s migrate
initContainer now also reads the config ConfigMap.
A game-type registry (server/src/tavolo/games.py) is now the single
source of truth for the games the platform can host; only scopone
scientifico is registered so far. GET /api/game-types exposes it for the
lobby's new game dropdown, and POST /api/games accepts a validated
game_type (default scopone_scientifico) which is carried on the live
GameState and onto each finished Match row (new indexed column,
migration 1_20260916235833_update), so statistics can be scoped per
game: /api/me/matches and /api/leaderboard take an optional game_type
filter and every serialized match includes its game_type.
Game states serialized before this change still load with the default
game type.
Granian serves the compiled SPA assets directly in Rust: hashed js/wasm/css
under /static (the release build uses --public-url /static/) and the card
images under /assets, configured with the GRANIAN_STATIC_PATH_ROUTE/MOUNT/
DIR_TO_FILE env vars in the Dockerfile. The Python catch-all now only serves
the SPA shell (index.html) at / and for client-side routes.
Every module logs through getLogger(__name__): lifecycle and business events
at INFO, per-move and store detail at DEBUG. The built-in default writes
DEBUG to the console; LOGGING_CONFIG points at a YAML file in the
logging.config.dictConfig schema to take over the configuration. PyYAML
becomes a direct dependency.
Single-file deployment under deploy/k8s/tavolo.yaml:
- Namespace, ConfigMap and Secret (placeholders) for the app
- Ephemeral Redis Deployment + Service (sessions and live games are
disposable, mirroring the docker-compose no-volume choice)
- App Deployment with an aerich-migrate initContainer so the schema is
upgraded before rollout, /api/health probes and a hardened pod
security context
- ClusterIP Service; Postgres and the OIDC provider stay in their own
namespaces and are referenced by DNS
The platform now hosts multiple card games, with scopone scientifico as
the first one. Rename the brand wherever it is not a game rule:
- move the Python package to server/src/tavolo and update imports
- rename the Postgres database/user, OIDC issuer path, client id and
Redis key prefixes to tavolo (clean break: existing pgdata volumes and
live games are not migrated)
- rename the Cargo package to tavolo-web and set the page title to Tavolo
- update docs and the Docker image path to woggioni/tavolo
The scopa game term (clearing the table) in the engine, state and web UI
is intentionally left untouched.
After each hand of an unfinished match the game now pauses in a new
hand_end phase instead of dealing immediately:
- engine: hand_points gains an 'award' map (which team won each category),
_end_hand stops at hand_end with a deadline, new acknowledge_hand deals
the next hand once all four players have acked; plays are rejected while
the summary is up
- state: acked seats, hand_end_deadline and hand_ack_timeout are persisted
and exposed in the personalized view (also on the finished state, so the
final hand is explained before the result)
- ws: new {"action": "ack"}; a per-hand timer force-deals the next hand
after HAND_ACK_TIMEOUT_SECONDS (new env var, default 30s) so an away
player cannot stall the match
- web: modal explaining each category in plain language with icons (card
images for denara/settebello/primiera), team-coloured rows, running
totals with progress bars, an 'Understood — next hand' button that turns
into 'Waiting for …' plus an auto-continue countdown; the final screen
shows the last hand's breakdown too
Verified in the browser against the compose stack: hand played to
completion, summary rendered (including a carte tie), ack from all four
players dealt the next hand live, and the auto-continue path fired when
nobody acked. 60 backend tests + mypy + cargo tests green.
Mirrors the reimpasto blueprint pipeline: tag release/x.y.z builds the
image (backend + WASM frontend in one multi-stage build) and pushes
gitea.woggioni.net/woggioni-opencode-agent/scopa:x.y.z plus :latest.
The build context is the repository root and the Dockerfile is
server/Dockerfile (monorepo layout). Requires the repo secret
PUBLISHER_TOKEN with write:package scope.
- Login/Logout anchors now carry rel="external": sycamore-router
intercepts same-origin anchor clicks otherwise, so the OIDC login
round-trip never left the SPA (found with a real browser test)
- Opponents' hidden hands render as a tidy 5x2 grid of card backs, seats
have a fixed height and shrink-wrap so the table no longer jumps as
cards are played
- Dockerfile: cache the Rust target dir so frontend rebuilds are fast
- ignore .playwright-mcp/ tooling output
woggioni