The news screen of SPEC 13 lists what the bot read in the last 24 hours and
what it did about each story, with the three filters of the mockup, the mood
of the market, the calendar with the effect of each event, the sources and
the spending of the day. Each note of the bot now keeps its action sentence
and the end of its restriction, so the screen shows the decision as it was.
Settings gain a news section: the rows of SPEC 10.5, the sources (enable and
trust), the key and daily cap of the language model, the Finnhub and FRED
keys and the central bank meetings. The demo seeds the F.9 data and matches
the mockup; the overview tile now shows the real news that matter.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Confirmed serious events hold new purchases of an instrument for 72 hours
and keep its stop where it is; confirmed negative news holds them for 24
hours; an unconfirmed rumour only gets a diary note and a check two hours
later. Coming earnings keep mean reversion out and macro figures hold
entries around their release, now in the live engine too. The strictest
row wins, every row can be switched off in the settings, and windows count
from the fact, so after an absence only effects still in force apply.
A held purchase moves to the end of its restriction while the order is
still valid and keeps its first due time, so the price drift check of
SPEC 5.6 still compares with the decision. Each restriction is written to
the diary once, with its motive. The per-instrument state keeps the tone
over one and three days, the story count against its norm and the active
restriction.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the restarts gone, a rerun still failed 7 tests on read timeouts. The
server log shows why: every DROP DATABASE of the tests forces an immediate
checkpoint, one at a time, and on the runner's disk the sync alone took up
to 37 seconds, so queued drops passed the client's 30-second timeout. The
service is a throwaway, so initdb now sets fsync, synchronous_commit and
full_page_writes off. docker-compose.yml and real databases keep them on.
ADR 0015 updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reverts the temporary three-round job. With --init and the shorter pg_isready,
three PostgreSQL suites on one server passed with no restart (run 321), where
the same run without them restarted the server four times (run 315). The
Docker events in the service log step stay.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The server log of a failing job shows every restart as 'untracked child
process exited with exit code 2'. A health check that outlives its 3-second
timeout leaves pg_isready to PostgreSQL, which is PID 1 in its container;
when pg_isready exits the postmaster takes it for a crashed backend and
restarts every connection. --init reaps such orphans, and pg_isready -t 5
under a 15-second timeout is never cut off. Same change in docker-compose.yml.
ADR 0015.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The news collector runs alongside the market data in the server test, so its
feeds and the Hugging Face download of the classifier can reach the fake network
before the assertion: they are the hosts the code declares, nothing else.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two clean runs did not show the restart. This collects the server log of a
failure sooner and lists Docker's events for the service container (oom, kill,
die). Reverted before the merge.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The classifier is distilroberta-finetuned-financial-news-sentiment-analysis in
Xenova's quantized ONNX export, pinned at revision e6953e30 and checked by size and
SHA-256.
- Model download: the model is not in the repository. The news job downloads it in
the background into the models folder of the data directory, the first time, and
retries after 6 hours if that fails. Until then the keywords judge alone, as
SPEC 10.3 requires.
- Tokenizer: RoBERTa's byte-level BPE travels inside the DLL, read with
BpeTokenizer and the RoBERTa pre-tokenizer. CodeGenTokenizer.Create adds ids for
runs of spaces that the model rejects. No special tokens are recognised in
untrusted text.
- Token check: the test compares the ids on 14 headlines with an independent copy
of the GPT-2 encoder run on the same files.
- Model test: it reads the tone of clear headlines at 23-24 ms per title on one
thread. It runs where the model is already downloaded and is skipped in CI.
Microsoft.ML.OnnxRuntime 1.31.1 and Microsoft.ML.Tokenizers 2.0.1 come from the
ADR 0004 list. Google.Protobuf joins the known indirect dependencies.
Anubis and Anubis.Server now declare win-x64 among their runtime identifiers, so
their lock files hold the Windows native packages. Publishing the Windows zip no
longer rewrites the lock files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The job log shows the tests losing every connection but not why the server
went into recovery. A new tasks command reads the service container's log,
state and memory through the Docker socket the runner mounts in the job.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This is the first judgement of SPEC 10.2 point 4, without a language model.
- Serious cases: cyberattack, suspension, fraud, bankruptcy, investigation and
delisting are found by whole-word phrases in English and Italian, so an
Italian headline still counts.
- Tone: the five tone grades follow the bounds of the specification.
- Importance: 4 for a serious case, 3 for a negative tone, 2 when the article is
linked to an instrument, otherwise 1.
- Classifier: it sits behind IToneClassifier and is asked only for English
titles. The ONNX model plugs in there; without it the tone stays unknown and
everything else works.
Every article stores its score, case and importance. Each story keeps the mean
score, the case and the highest importance.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
News is the first input of Phase 7. Every two minutes the "notizie" job asks each
enabled feed conditionally and backs off a failing one up to an hour.
- Parsing: RSS 2.0, RSS 1.0 and Atom are read without DTDs or entity expansion, so a
hostile feed reads nothing from disk.
- Deduplication: articles are deduplicated by canonical URL.
- Grouping: retold titles join one story by word similarity within 48 hours, and
independent publishers are counted by registrable domain.
- Linking: stories are linked to instruments by source symbols, cashtags, exchange
tickers, coin tickers and capitalised company names; to the crypto or equity class
or to the market by theme; to the source topic as a fallback.
ADR 0014 records the thresholds. The tests read a synthetic archive in each source's
real format and run on both engines.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Trying the packaged release showed three things to fix: a 101 answer that opens a
WebSocket was counted as a failure, the duration buckets were too coarse around
the usual 200-500 ms of the brokers, and the report mixed CRLF and LF lines and
named SQLite only by its version.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 48-hour soak run at home kept being cut short by the host, never by Anubis.
The user will instead run a release for as long as they like and paste back a
report, so the bot now records what that report needs (ADR 0013).
- Every 5 minutes and at each start, a process sample goes to diagnostic_samples:
memory, GC, threads, handles, CPU, exceptions, database size, log warnings and
errors, and outbound HTTP calls by destination with duration buckets, read from
the System.Net.Http meter without any new package.
- Every scheduled job run goes to job_runs. Both tables keep 30 days.
- The Markdown report (at most 60 KB, redacted, user folders masked) adds
findings, the soak checks, starts and silences, jobs, calls, grouped log
warnings and errors, and resources over time with their slopes. It comes from
"anubis diagnostics", GET /api/diagnostica.md and a button in Diario e sistema.
- New {long} column token: SQLite STRICT tables reject BIGINT.
- "tasks release-win" builds the trial zip with a start script, a script that
writes the report, and Italian instructions. Version 0.7.0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On 12 November at 23:37 Docker Desktop stopped for 15 minutes. It had
been started from the Claude app's shell after the PC switched off, so it
lived in the app's process job and died when the app updated itself. The
bot resumed alone and recorded the absence.
The run is not restarted from zero: a fresh start would no longer span
two US sessions before the weekend. Instead two reports are read: the 48
uninterrupted hours from 23:57, which must be regular, and the whole run
from 21:02, whose only allowed failure is that 15-minute stop. Docker is
now started through WMI, outside the app, and SVILUPPO.md says how.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first attempt ran 22 hours: on 12 November at 19:49 the PC switched
off without a clean shutdown (power or a machine hang, no bug check).
When it came back the bot resumed by itself and wrote the absence in
the diary, but a 48-hour run needs no gaps, so it started again from
zero at 21:02. STATE gives the new window and command; the first
attempt's report and database are kept under artifacts/prova-48h.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The run started on 11 November 2026 at 21:44 Italian time in the
anubis-prova container, from the image of this branch. STATE says when
and how to make the report so that it spans both US sessions; its check
is the first step of Phase 7 (#55).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The scheduler's first jobs finish about a minute after the bot starts, so
a run started exactly at the window's start was reported as covering
slightly less than 48 hours. A gap shorter than the allowed silence
(15 minutes) at either edge now counts as covered, like one in the middle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Starting the 48-hour run showed two defects.
After an absence, a decision made past 12:00 UTC was written off as
"no data" at the first pulse, because the data step had not yet brought
the bars of the close: the recovered decision was lost. A late decision
now asks for its bars and waits up to 30 minutes for them. The 20-hour
absence test pulses before the bar arrives and fails without the fix.
The soak report measured coverage and silences from reconciliation
rows, but a clean round writes none, so a healthy run would have been
reported as never covered. It now reads when the scheduled jobs first
ran and last finished, the absences written at each restart, the guard
interventions, and each account's status (reconciled within the last
15 minutes and normal). The test no longer fakes a row every 5 minutes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An independent review of the phase found nine defects. The ones that
could lose money or orders:
- Exits at the close of equity_meanrev are now planned orders written in
the same transaction that claims the decision, so a restart can neither
lose them nor send them twice. A sell with a deadline starts aggressive.
- Resume also finds a planned order not yet linked to its execution
(restart between sending and linking) and closes it before re-planning.
- An account is "ready" only after adopting the guard's stops, checking
the key, reconciling, resuming and aligning stops. Until then nothing is
decided on it or sent to it, and the loop retries its startup each
minute. The guard's note is deleted only when every stop it placed is in
the ledger.
- Each account's turn is cut after 45 s (60 s for a retried startup), so a
slow broker no longer silences the heartbeat until the guard kills a
healthy engine; the beat is also withheld when the scheduler has been
silent for 30 minutes.
Smaller ones: a stock decision made after the open is marked late; the
Telegram loops survive unexpected errors; a long absence is shown on the
Panoramica until "Ho visto" (SPEC 5.6); plain-http heartbeats are only
accepted to home-network addresses; the soak report gives a daily close
30 minutes to be decided before expecting it.
ADR 0012 point 12 records these choices; #51 and the two screen criteria
are ticked in the roadmap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"anubis soak" reads the bot's records over a time window and says whether
the run kept control. It changes nothing, so it can run beside the working
bot with docker exec. It exits with 4 when something needs a look. It
checks:
- the hours covered and silences longer than 15 minutes;
- orders or decisions made twice;
- unexplained reconciliation differences and each account's last outcome;
- absences and guard interventions;
- one crypto decision per daily close.
ADR 0012 records what the specification leaves open in this phase: the
live engine shares the backtest's decision call, run keys, recovery after
an absence, the guard and its drill seams, the diary templates, the
"senza bot" comparison, Telegram and the heartbeat, the screens and the
endurance run. ROADMAP ticks the Phase 6 items and criteria with the tests
that prove them, STATE moves on to Phase 7, and SVILUPPO explains the
guard and the report.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The guard drill failed once under load. The pipe of the killed engine
delivered a late "running" beat after the restart. The drill took the new
engine as ready and stopped it before it had cancelled the half-done sale
and put the stop back.
The same late message could have made a crash of the new engine look like
a planned stop, which the guard never restarts. Each engine the guard starts
now has a number, and messages from a replaced one are ignored. The drill
also waits until the restart's work is in the database before it checks.
The morning Telegram summary listed every job due today, including the
background ones that run every minute. It now lists only the bot's moves.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Panoramica now follows mockup 01. Below "Finora" and "Adesso e dopo" it
shows:
- the two account cards with the invested/cash bar and a 30-day trend;
- the four risk limits;
- the open positions by value;
- the strategies;
- the news, learning and system tiles.
On the phone it follows mockup 09: Conti as a list, four positions with
"Vedi le altre", and two limits. Outside the demo, news and learning say
they are not active yet (Phases 7 and 8). The next moves are the scheduler's
moves only, no longer every interval job.
Diario e sistema follows mockup 08:
- the diary grouped by day, with search and filter chips (plain links, no
JavaScript) and older days on request;
- how each component is doing;
- the scheduled jobs;
- the host's processor, memory and disk;
- the technical logs as a zip, for signed-in users only.
In demo mode the screens show the values of appendix F.10. Tests cover the
demo figures, the filters and search, and the logs download.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every notice is written to the queue in the same transaction as what it
tells, then sent every ten seconds. Each has a state: pending, sent, or
failed with its error. Network failures and 429s are retried with growing
pauses (Telegram's own wait when it gives one). Refusals are final. A notice
that could not leave within a day is given up instead of arriving late. Every
fill now also reaches the phone, and so do pause, resume and restart.
Commands are read by long polling and obeyed only from the configured chat
and only when fresh, so yesterday's /pausa never pauses today. They are
/stato, /posizioni, /oggi, /pausa, /riprendi, /ferma (confirmed with
/conferma within two minutes) and /aiuto. Restarting after "Ferma tutto",
real money, limits and keys stay in the interface.
Also added:
- the heartbeat call every minute when an address is configured;
- the morning summary at 08:00, caught up within four hours;
- an optional hourly or evening status;
- a Telegram section in Impostazioni: test-before-save, a test button, the
queue's state and the last successful heartbeat;
- on desktop, the first-run page recommends the heartbeat.
Tests run against a fake Telegram server, including the commands through
the real server.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The guard (SPEC 9.3, 9.4) is now the container's main process and the parent
of the engine. It listens on a named pipe for the engine's heartbeat and
startup steps, and steps in after two minutes of silence, ten minutes of a
startup that does not move, or an exit without notice. It stops the engine,
cancels our purchases and sales left open on demo accounts, puts back the
stops the positions need from the state file the engine writes every ten
seconds, writes a note, warns the user and restarts the engine without any
injected fault. The third forced restart within an hour writes the STOP
file. The container's stop is passed on to the engine, which closes in order.
At restart the engine adopts the guard's stops before reconciling, cancels
the orders it left open and plans again the exits already decided, so the
rules of SPEC 5.6 apply to them.
The acceptance test blocks the main loop right after a sale leaves: in
process on a demo-like broker, and with real processes in the drill, which
also runs inside the Docker image (tasks ci and the CI workflow). The guard
uses the ASP.NET shared runtime so both programs share /app.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
At the start the host measures how long the bot was away from the last
job that finished, records the absence and writes it in the diary; one
longer than a day is kept for the Panoramica until the user has seen it.
A job can now say how a run missed while the bot was off is treated: the
nightly backup, the exchange rates, the market calendar and the cleanup
run once at the restart, queued, however late; a job with a deadline is
skipped past it and the diary says so. Corporate actions are read before
the reconciliation puts any stop back. A decision the engine makes late
because of an absence is written in the diary as recovered, as are the
orders it releases late; entries that moved too far or expired are
dropped with their reason.
Tests switch the bot off for 3 hours, 20 hours and 3 days with a manual
clock and check each of those rows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The live engine makes each decision at its time of SPEC 5.5 through the
same DecideOnce call the backtest uses, so live and backtest cannot drift:
the crypto close at 00:05 UTC, the stock signals 20 minutes after the
close, the meanrev exits 5 minutes before it, the stock orders and the
daily stops at the open. Each decision has a run key written in the same
transaction as its decision rows, planned orders and diary lines, and each
planned order's id gives the order key, so a restart never decides or
sends anything twice. Entries recovered late are dropped when the price
has moved too far.
The engine runs from the scheduler once the accounts are reconciled. The
IEX stream follows held stocks, planned orders and SPY during the session,
and the stock simulator's day orders now expire at the close.
Tests: parity with the backtest for crypto_trend (40 days) and
equity_meanrev (6 weeks), and a 14-day run with the network down, a lost
answer, a restart mid-order and the database unreachable, ending
reconciled with no duplicate orders, on SQLite and PostgreSQL.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The diary needs sentences in the tone of appendix B, built only from what
happened. DiaryText holds the templates the live engine will use: fills
with the result of the trade they close, raised stops, the daily crypto
close, the stock signals, observation signals, recovery and absence. The
reason-code sentences move from the server into the library
(ReasonCatalog) so the engine and the screens share them; a test checks
that every reason constant has its sentence. The diary gets the kinds of
the Diario screen: Ordini, Decisioni, Sistema, Notizie, Apprendimento.
The comparison of SPEC 7.5 invests each capital flow on its day, 40% in
BTC/EUR and 60% in SPY adjusted and converted at the ECB rate, and never
touches it; a withdrawal sells both in the proportion they reached. Prices
are the closes of the day before, so a deposit is never a gain. Without
the SPY history (no Alpaca keys) the photograph leaves it empty rather
than show half a comparison. The Panoramica now takes the flows out of
the "Senza bot" result, as it already did for the bot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Loading the outcomes before "ready" was not enough: on a fresh database the
strategy rows did not exist yet, so the load did nothing and CI run 220
failed the same way on PostgreSQL. The new TradingService.PrepareAsync
creates accounts and strategies, then loads the outcomes; the host runs it
before it is ready and logs a failure without blocking the interface.
StartAsync repeats it, which changes nothing.
The demo tests deleted the demo folder while the server still held its
SQLite file, which Windows refuses at random. They now stop the server
first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The trading service loaded the outcomes in the background after the host
reported "ready", so for a moment Strategie showed every strategy as
without a backtest. CI run 219 caught it on SQLite. The host now loads
them before it is ready; the call is idempotent, so the one at trading
start changes nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An independent review found these, now fixed and tested:
- Risk limits could be raised through PUT /api/settings without the
phrase and the password. The API now refuses risk keys.
- The reference equity counted results twice after reduced to real and
after a demotion. Each ledger now keeps a baseline, and carried results
stay in the ledger's currency until the reference is computed.
- An instrument could be bought by a second strategy in the same account
(SPEC 8.6), and purchases under way did not hold their cash, so several
purchases sent together could exceed it. Both are refused or shrunk now.
- After an automatic trip of the switch there was no way to restart from
the interface. Rischio now shows the trip of both worlds and the restart.
- The fresh password of SPEC 12.4 could be tried without limit. Five wrong
attempts in 15 minutes now lock it.
- Observation to simulation skipped the 90 days of live signals.
- A new build changed the qualification of a strategy with real money,
"Ferma tutto" counted as a critical incident, raising risk ignored
demotion conditions in progress, and weekends counted as stock sessions.
- The 12.4 confirmation now summarises capital, first purchases and limits,
and live strategies show the gap with their simulated twin.
The backtest converts each bar with the rate fixed before it opens, the
day before, and the report now counts how often correctives and risk
changed a decision. crypto_trend is still not passed: the broad universe
now falls 32.8%. The outcome file also keeps the qualified parameters.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ADR 0011 collects what the specification leaves open or refines in this
phase: strategies shared by backtest and live, correctives and risk inside
the backtests, the risk engine that only refuses or shrinks, the stages as
the only way to real money, the euro history and the protocol of the base
backtests, and the honest "not passed" of crypto_trend.
ROADMAP ticks the Phase 5 items and criteria with the tests that prove
them; STATE moves to Phase 6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The three screens follow mockups 03, 04 and 07.
Operazioni shows the balance of the closed trades, filters, a CSV export,
and for each trade why it was bought and sold. The reasons come from the
decision codes through fixed Italian templates.
Strategie shows the four stages towards real money, a card per strategy
with its requirements, the procedure of SPEC 12.4, and why each strategy in
observation waits. A strategy that failed its backtest cannot be moved into
the simulation.
Rischio shows the emergency switch, the automatic blocks, the use of each
limit, the fall from the peak, the scenarios and the risk events. Stop,
close and restart now go through the risk engine. A restart is refused
while the STOP file exists.
Impostazioni gains the limits, the day-trade guard and the scenarios.
Lowering a limit is always allowed. Raising a limit or the entrusted capital
while a strategy uses real money needs the phrase and the password.
The demo now holds the decisions behind its trades, the 58 days of
simulation and the indices of its curves, so it matches the mockups.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crypto_trend now has a backtest report and an outcome file, computed by
"anubis backtest base" on the Binance daily archives from 2019 to October
2026. Prices are converted from USDT to euro with the ECB rate of the day,
and delisted coins stay in the universe, so survivors do not flatter the
result. The protocol runs the broad universe and BTC, ETH and SOL, with
costs doubled and tripled and buy and hold beside. Both universes must pass.
The outcome is "not passed": the broad universe falls 29.1% (limit 25%) and
the narrow one has a probability of overfitting of 0.84 (limit 0.40). The
strategy therefore stays in observation. The stock strategies wait for the
Alpaca keys and stay in observation too. The bot reads the outcome files at
every start.
Trials now read 450 days of history before their start, only so the
indicators and the 400-day listing age are ready on the first day. Before
this, 2019 was spent warming up and BTC and ETH could not be chosen until
2020. A daily series also stops at a hole longer than a week, so the LUNA
relaunched in 2022 is not mistaken for the coin that collapsed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every intent of a strategy now crosses a risk engine that only refuses or
shrinks: loss limits per world with the 24-hour block, the weekly halving,
the halving and the switch on the fall from the high, the size limits on
invested capital, single position, crypto share and trade risk, the
technical checks, no short sales and no leverage, the day-trade guard
(off by default). The switch, "Ferma tutto", "Ferma e chiudi", the STOP
file and the orders-per-minute anomaly stop the bot until the user
confirms a restart, and a restart after a trip measures the high from
then on.
The correctives of SPEC 8.5 (crypto regime, economic calendar) and the
risk rules also run inside the backtests of the base strategies, so the
reports show what would happen live. Each stage of a decision is written
to a new decisions table with reason codes.
The stage book is the only way to real money: an order on a real account
needs the startup consent, every requirement of SPEC 12.3 met, the typed
phrase and a fresh password and second factor. Order managers on real
accounts fail closed and ask again before each order leaves. A demoted or
revoked strategy keeps its real positions in exit only, with stops, until
they are at zero. The simulated twin stays out of totals and of the
reference equity.
Tests cover each rule of 9.1 and 9.2, with property tests for "never
larger than asked" and "never larger after a loss", and both proofs on a
fake real account: nothing reaches it without the procedure, and a
demoted strategy manages exits and stops down to zero.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crypto_trend follows nine breakout periods with Donchian-middle stops,
volatility-targeted weights under the 35%/20% caps and a monthly top-10
universe. equity_momentum reviews on the last close of each week with the
SPY filter, inverse-volatility weights and an ATR trailing stop.
equity_meanrev buys RSI(2) dips above the 200-session average, sized for
a fixed risk, and exits at the close.
The engine gains what the rules need: exits at the same close, a stop set
at a distance from the price obtained, data-only instruments such as SPY,
the trading calendar, the reference equity of the strategy's share and the
equity history. Every strategy passes the look-ahead check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Four launch entries for the server: a data folder of its own under
artifacts/ (ignored by git), the demo mode, the user's real data folder,
and attach to a running process. Each builds in Debug first and opens the
browser once the server listens. tasks.json also exposes check, test and
dev. SVILUPPO.md explains which entry to pick.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- An error of the database (PostgreSQL restarting) no longer stops the
laboratory until the next start: the queue logs it, waits 30 seconds
on the injectable clock and tries the same trial again.
- Buy and hold, the comparison of every report, buys an instrument
whose history starts later at its first close, with the share kept
free for it; before, it was never bought and the others took its
weight. Its version goes to 2.
- The comparison page reads a and b as text: a malformed id shows
"two finished tests needed" instead of an error page.
- A trial re-checks the hash of its plan before running and refuses a
plan that no longer matches it.
- Parameters that count days take whole numbers only (10.5 would run
as 10 and count twice in the DSR); the capital starts from 1, as the
message says.
ADR 0010 says so, and corrects the estimated minimum of the research
history: 5 does refuse smaller orders, like the exchanges' minimums.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ADR 0010 writes down what the specification leaves open: the event
order of a session, the costs shared with the simulator, the pass-through
correctives and risk until Phase 5, the in-house statistics, validations
on the daily returns of the variants, the plateau rule as the minimum of
own and neighbourhood Sharpe, the sizes of walk-forward, CPCV and CSCV,
the trial count of the DSR, the plan written first and never changed,
the low-priority laboratory and the two data origins (the research
history stays in USDT; a trial uses a single currency).
ROADMAP ticks the four acceptance criteria with the tests that prove
them; STATE moves to Phase 5.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The laboratory starts with the host (off in demo mode) and runs its
queue next to market data and trading. The page lists every trial with
status, verdict and headline figures, refreshing itself while one is
queued or running, and offers the form for a new test: strategy, data
origin and market, instruments, period, initial capital and the grid,
with the thresholds stated before the run. The form posts with the
antiforgery token; errors come back as a short code.
The report page shows the protocol checks, the chosen variant against
buy and hold (table and chart), results per year, cost sensitivity,
the Deflated Sharpe Ratio, PBO, walk-forward windows, purged CV paths,
the look-ahead result, every variant with its zone score, what the
backtest does not model, the data used and the trades. A comparison
page puts two finished tests side by side and says whether they read
the same data. The improvements of 8.7 and the cost estimates of 11.6
have their sections, empty until Phases 9 and 8.
The equity chart takes amounts in their own currency and its labels as
parameters. The profit factor has no value without a losing trade.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A trial is planned before it runs: strategy and version, grid, data,
initial capital, cost multipliers, validation settings, thresholds and
engine version go into backtest_trials as JSON with the SHA-256 of the
text. The store has no way to change a plan; later writes touch only
status, verdict, results and times, and a test reads the raw row back
to prove it. Every finished trial stays and counts towards the number
of tries the Deflated Sharpe Ratio deflates for.
The trial runner is pure computation: every grid variant on the whole
period, plateau choice, the chosen variant again at double and triple
costs, buy and hold, walk-forward, purged CV, PBO, DSR and the
look-ahead check, then the five thresholds of SPEC 12.2 and the report
of point 6 with the limitations the backtest declares (no correctives
without history, no risk engine before Phase 5).
The laboratory writes the plan, queues the trial and runs one at a time
on a dedicated thread with below-normal priority; only the async side
reads and writes the database. On start, trials left running go back
to planned and run again. Data comes from the daily bars and corporate
actions in the database, or from the Binance research history with
estimated order rules.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Metrics give what the report of SPEC 12.2 point 6 needs: return, CAGR,
annualised volatility and Sharpe, max drawdown with its days, trades,
win rate, profit factor, fees, exposure and results per calendar year.
Statistics are written in house as SPEC 12.2 asks: the normal CDF
(Hart 5666, double precision) and quantile (Acklam plus a Halley step),
moments, the probabilistic and the Deflated Sharpe Ratio. The DSR test
reproduces the example of Bailey and López de Prado (SR0 0.1132, DSR
0.9004).
Validation works on the daily returns of every grid variant:
walk-forward with a growing window, combinatorial purged CV (6 groups,
2 tested: 15 splits, 5 paths, purge and embargo of 1% of the series)
and PBO by CSCV (16 groups). Pure noise gives a PBO near 0.5, a
dominating variant 0, a winner that always loses out of sample 1.
Choices follow the plateau rule: a variant scores its own Sharpe but no
more than the mean of its zone, so a lonely peak loses to a broad hill
and a poor variant does not win by standing next to a good one.
The look-ahead check reruns the strategy on data cut at six days and
requires every decision up to each cut to be unchanged; it fails on a
strategy that reads tomorrow and passes on the average cross. The
one-step comparison of SPEC 12.3 classifies intents and flags action
differences and sizes more than 10% apart. A golden run guards the
engine against unintended changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The strategy contract of SPEC 5.3 (IStrategy, StrategyContext, Intent) is
shared by backtest and live: the context exposes bars only up to the
decision, adjusted and anchored to that day, and reading past it throws.
The engine runs each session as ordered events (corporate actions, open,
session, close), decides on the close and executes at the next open with
the simulator's price lists, slippage tiers, tick and lot rounding,
minimums and no leverage. Corrective and risk hooks are wired in and pass
everything through until Phase 5. One decision cycle can restart from a
recorded book (SPEC 12.3).
A scripted toy strategy reproduces a result computed by hand to the cent
on OKX and Alpaca costs. The Phase 3 class Trading.Strategies becomes
StrategyNames so the Anubis.Strategies namespace can exist.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI run 188 failed once on the SQLite job: a cross-engine test lost its
connection (EndOfStreamException) while creating its PostgreSQL database,
right after the service container started. The retry covered only
timeouts. It now covers every transient Npgsql error and picks a new
database name on each attempt, in case the lost answer had already
created the database. Test helper only; no product code changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Splits multiply only the shares held before the ex-date. Old splits in
the table no longer inflate positions bought later; the simulator's book
follows the same split once, together with the ledger.
- Dividends pay the simulator once on the ex-date holdings, and each
payment is marked in the same write, so a failure never pays twice.
- Positions are marked to market before every equity photograph; until
now equity, drawdown and the Conti screen stayed at the last fill.
- A sale with no price keeps its broker stops until it can really leave;
a purchase with no price is dropped.
- A refused sale is retried with pauses of 1, 2, 4, 8 and 15 minutes
instead of every tick, and one failing execution no longer stops the
stop alignment of the whole account.
- On Alpaca a purchase clears the sell stops of every strategy on the
symbol and holds them until it ends.
- OKX order lookups treat only 51603 as unknown; any other failure is an
error, so a lost order is never sent twice.
- The reconcile choice and key removal forms now require the antiforgery
token; the choice re-runs the comparison and is only offered after the
15 minutes; stale differences are dropped.
- A replaced key that cannot withdraw releases the prudent mode.
- Re-entering the same euro capital for Alpaca changes nothing; a new
amount converts only the euro difference at today's rate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ADR 0009 records the Phase 3 choices; API-NOTES gains the verified trading
facts and what Anubis does with each; ROADMAP and STATE mark the phase done
with the live-demo conformance waiting for the user's keys.
Two defects found while writing the notes:
- The positions CSV used dot decimals; SPEC 13 asks for decimal commas.
Trailing zeros are dropped so both engines give the same text.
- Alpaca's catalogue has no tick, so limit and stop prices were sent
unrounded and would get 422 "sub-penny". OrderRules.TickOf applies
Alpaca's rule (2 decimals from 1 $, 4 below) to orders, stops and the
simulator's own check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Conti e posizioni screen follows mockup 02: totals split by strategy,
one card per account with key permissions and the last reconciliation,
positions with a per-account filter and CSV export, pending stops (Alpaca
grouped into permanent and day stops), residues, and exposure by currency.
It shows one world only, real accounts when there are any, so simulated
and real money never add up. Position values are rounded to the cent per
row, as the table shows them, so the figures add up on screen.
Settings gain "Conti e chiavi": practice keys are asked to the broker
before they are saved and refused when they may withdraw or cannot trade;
they are never shown again and the engine restarts to use them. The
capital entrusted to each broker is set there in euro. Until any capital
is set, the overview points to that section as the first step.
The overview now reads one world, nets capital flows out of its results
and uses the time-weighted index; open orders exclude protective stops.
The demo seed gains cash rows, closed trades, stops, the ECB rate and the
reconciliation state, so the ledger itself computes the mockup's figures.
Screenshots now capture the whole page and include Conti.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One abstract battery states the adapter contract (time and cash, an order's
client id from send to cancel, well-formed orders, positions and fills with
fees as costs, key permissions) and runs it on the internal simulator, on
Alpaca's and OKX's documentation samples with stateful fakes, and on the demo
accounts when the user's keys are in the environment (LiveDemo, inconclusive
without them). The OKX sample key lists withdraw, and the adapter reports it.
Service tests cover the accounts created without keys (no request leaves),
stock strategies moving to the paper account only while their ledger is
empty, the simulator funded from the ledgers once, and dividends and splits
applied once each. The whole suite also passes on PostgreSQL.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Order manager tests drive the internal simulator turn by turn: the chasing
limit with its post-only steps and aggressive cross, a buy given up, a sale
never given up with its alert, the pace of six sends a minute, Alpaca's
GTC + day stops, the stop-already-passed rule and splits, lost answers,
duplicate fills, refusals and residues. Reconciliation tests cover clean,
healed, prudent mode and its return, align/exclude, unknown orders, stop
coverage, withdraw-capable keys and an unreachable broker.
Two defects found by the tests are fixed: stops were not cleared before an
execution that starts aggressive (Alpaca refuses a buy with a sell stop open,
and a sale must free the quantity), and an order's creation is now recorded
as its first event.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 3 needs books the bot can trust before any strategy trades. This adds:
- schema 0003: executions, order events, trades, residues, account status,
reconciliations, capital changes, applied corporate actions, simulator books;
- the broker interface with the internal simulator and the Alpaca paper and
OKX demo adapters, all behind the live guard (SPEC 2.2);
- the append-only ledger with lots, trades from zero to zero, residues,
dividends, splits and reversals; capital entrusted per broker and equity
photographs with time-weighted return and FX effect;
- the order manager (chasing limit, pacing, sells never abandoned, server
stops, breached-stop rule) and the reconciler with prudent mode;
- the trading service wired into the host, off in demo mode and in tests.
Ids made with one clock in the same millisecond now carry a counter (RFC 9562
method 1): two lots of one order kept a random order, which broke LIFO.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On the CI runner twelve test workers open PostgreSQL connections at once
and some waited longer than the application's 5 seconds, failing two
collector tests with a connection timeout that has nothing to do with
the code under test. The connect timeout is now a DbSettings field that
keeps the 5-second default for the application, and the test databases
ask for 30 seconds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The skipped-minute test waited for the recovered bar, but the issue is
closed a moment after the bar is stored; on PostgreSQL in CI (run 181) it
read the issue in between and saw it still open. It now waits for the
issue's final state.
- Creating a PostgreSQL test database retries on a connection timeout: CI
run 179 on main failed because twelve workers created their databases at
once and one connection missed the application's 5-second timeout. The
application's timeout is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Gaps between the last stored bar and the first new one were never found,
so a minute the live feed skipped, or a day missing right after the stored
history, was lost for good. A gap is now new when it ends after the last
bar held, and is downloaded again like any other.
- When one OKX connection dropped while the other kept writing, the write to
the closed feed threw out of the live loop and left it marked connected.
Writes now stop quietly, and the loop treats any failure (database
included) as a drop: marked down, back off, reconnect.
- The background worker now survives a failed round instead of stopping for
the life of the process.
- Alpaca gives adjusted prices on the basis of the download day, so a later
split looked like a crash. Each download re-reads the last stored day and,
if its adjusted/real ratio changed, rescales all older adjusted prices.
- Stock days are fetched up to the last closed New York session, so the job
after the close collects that day; missing stock days are checked against
the exchange calendar; delisted stocks with history are not asked again.
- Smaller: a lock in the freshness monitor, no NaN mood readings (SQLite
stored them as NULL), and the network proof runs only with
ANUBIS_LIVE_PROOF=1. The test wait is 30 s for busy machines.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ADR 0008 collects the market data decisions the spec leaves open: research
history in csv.gz instead of Parquet (no extra package), the immediate second
reading of suspect data, EUR pairs only, Alpaca only with the paper keys,
name-based fund exclusion and the background worker. API-NOTES gets the facts
checked while writing the adapters and the biases that remain in past
universes. ROADMAP and STATE mark the phase done, with the stock history still
waiting for the user's Alpaca keys.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A LiveDemo test, outside the normal suite, that downloads the whole OKX EUR
daily history of BTC, ETH and SOL and the Binance BTCUSDT archives since 2017,
then writes the quality report. It is the evidence for the phase criteria that
a fake network cannot give.
The detached database test now retries the file move for a few seconds:
Windows refuses to move the SQLite file while a request of the running server
holds it, which made the test fail now and then.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A "Dati" section lists every external source with its state, last answer
and the reason it is off (Alpaca without the paper keys, FRED and Finnhub
without theirs), links the 30-day quality report, and edits the Fed/ECB
meeting list; a list that does not parse is refused with an example.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MarketDataService wires the collector, the reference sources, the
universes and the corporate actions into the scheduler. Short jobs (rates,
calendar, mood, spread samples, corporate actions) run on their schedule;
long downloads (catalogues, the first history, thousands of stocks, the
Binance archives) run on a background worker the jobs only wake, so they
never delay the backup or each other. The live crypto feed follows the
crypto universe, or BTC, ETH and SOL until one exists.
Alpaca uses only the paper account keys; without them the stock part is off
and the sources list says why. Market data is off in demo mode, with
ANUBIS_MARKET_DATA=off, and in tests that bring no fake network.
New: GET /api/data/quality (JSON or Markdown) and /api/data/sources, and
"anubis data report|binance|csv" on the command line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MarketCollector downloads what is missing at startup and after every drop
of the live feed, checks every bar and never corrects in silence: suspect
bars are stored as suspect and read a second time (same values confirm,
different ones replace, impossible ones are dropped), gaps are downloaded
again once and what the source still lacks is recorded as unrecoverable.
A test detaches the feed for ten minutes and finds every minute recovered.
The stock and crypto universes are chosen "as of a date" from the bars up to
that date only: funds, ETFs and ETNs are excluded through nasdaqtraded.txt
and the security name, stocks delisted later still count when they traded.
Corporate actions are kept once per source key and found by ex-date.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Readers for the ECB rates, the Forex Factory calendar, FRED release dates,
Finnhub earnings, the Fed/ECB meeting list, fear and greed, the Cboe VIX
history, nasdaqtraded.txt and the Binance kline archives, each tested on
recorded responses. ReferenceSources downloads them within their limits
and records every attempt in source_state, so limits hold across restarts
and the interface can say which source is off and why (FRED and Finnhub
without a key). The central-bank list is an editable, validated setting.
BinanceImporter checks every monthly zip against its SHA-256 file before
writing it to the research history; CsvImporter reads user files with a
configurable column mapping and reports unreadable lines and gaps instead
of fixing them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both clients are written in house on HttpClient and ClientWebSocket, behind
IMarketDataSource, with a rate limiter on the injected clock and retries
that wait on it too.
OKX: EUR pairs only (trap 14), rules from public/instruments, daily
candles with bar=1Dutc and a refusal of any candle not opening at 00:00
UTC (trap 23), backwards paging of history-candles, errors read from the
code field because OKX answers them with HTTP 200, tickers on the public
stream and candles on the business stream with ping after silence. The
HMAC signer and the demo header are ready for Phase 3.
Alpaca: consolidated daily bars, real and adjusted, dated in New York and
never newer than 15 minutes (trap 8), an explicit feed on every request,
IEX for recent minutes, quotes and the stream (at most 30 symbols),
calendar and clock converted to UTC, corporate actions. Without keys
nothing is sent and the reason is stated. Tests use the documentation's
examples and synthetic answers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Migration 0002 adds the market data tables on both engines: daily and
one-minute bars, corporate actions, FX rates, the exchange calendar,
calendar events, mood indicators, spread samples, data issues, source
state and universe members, plus the trading rules on instruments.
The data layer is format-independent so the venue clients can build on
it: the instrument catalogue (rules from the venue, delisted kept), bar
upserts with ON CONFLICT, the SPEC 6.4 checks with an issue register
that blocks entries until a second reading, freshness by age of the last
datum (trap 10), research history in .csv.gz with exact decimals, and
adjusted series anchored at the decision day with level and quantity
conversion across splits and dividends.
Recorded API responses from OKX, the Alpaca docs and the free sources
are kept byte for byte as test fixtures; the secret scan now ignores
pagination cursors such as Alpaca's next_page_token.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user allowed the download, so the interface now uses the typeface
of the mockups instead of the system one. The variable file (weights
400-700) and its SIL OFL licence are embedded and served by the app
itself, never fetched from the internet; source and SHA-256 are in
API-NOTES. The typeface is slightly wider, so top-bar labels and
buttons move to 13 px as in the mockup, keeping the date line whole at
1440 px.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An error-page choice released the startup loop before resetting the
services, so the loop could start the application and then lose its
services while reporting ready. The reset now happens while the loop is
still held, and moves the status back to Starting.
A SQLite file could vanish between the existence check and the open;
with create allowed, the open would put an empty database in its place.
Once the file has been seen, connections open read-write only and fail
on a missing file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 1 completed on top of the library foundations:
- Server: startup state machine (config, first run, database down, schema,
identity, ready) with an error page that retries by itself; the configured
database is the only one ever opened, and a vanished SQLite file is never
recreated. REST API and status WebSocket; session cookie auth, CSRF on
every form and write call; desktop and demo answer only the computer itself.
- Interface: static Blazor rendering with plain forms (no interactive
runtime), menu, top bar and phone bottom bar as in the mockups, light and
dark themes with appendix C tokens and a contrast test, Panoramica, Diario,
Rischio, Impostazioni; service pages for sign-in, first access, desktop
first run and the database error page with its choices.
- Demo mode (tasks seed-demo): appendix F in a database of its own under the
temp folder, frozen clock, never touches the configured installation.
- Docker: entrypoint gives /data to PUID:PGID and drops root with setpriv,
HEALTHCHECK through bash /dev/tcp; proven with SQLite, PostgreSQL, a wrong
configuration and a database that goes away and comes back.
- tasks screenshots drives the installed Edge/Chrome through DevTools for
exact 1440 and 390 pixel shots; anubis db export|import on the command line.
Fixes from the independent review: return paths without control characters,
Host check against DNS rebinding, attempt limit counted before the check and
also on the error-page password, one-use TOTP codes and a code to switch the
second factor off, error-page choices serialized with the startup loop,
identity checked before migrating, restore restarts with "Ferma tutto" on,
damaged archives reported instead of crashing. Running the server tests on
PostgreSQL exposed pages rendered before their data arrived; data-loading
components now load first (LoadedComponent).
Decisions in ADR 0007. Tests: 100 on SQLite and on PostgreSQL.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 1 foundations in the Anubis library:
- Platform: startup configuration from environment variables (Docker) or
config.toml (desktop, in-house TOML subset), data folder layout, secret key
custody (ANUBIS_SECRET_KEY, DPAPI on Windows, restricted file elsewhere),
installation identity and database fingerprint.
- Db: ADO.NET with Npgsql and Microsoft.Data.Sqlite, money as NUMERIC(38,18)
or canonical text rounded to 18 decimals, UTC moments, UUIDv7 ids; numbered
SQL migrations with type tokens and a version table; engine-independent
export/import with count and accounting-total checks; backups keeping 14.
- Core: injectable clock, channel event bus, scheduler with fixed, interval
and market-relative times, settings with change log, AES-GCM secrets, log
redaction, Italian formats.
- Access: PBKDF2-SHA512 hashes, RFC 6238 TOTP, hashed session tokens,
sign-in attempt limit.
SQLitePCLRaw (Apache-2.0) arrives with Microsoft.Data.Sqlite and is declared
in DependencyPolicyTests and API-NOTES.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The independent review found that the policy tests could be bypassed:
Microsoft-named direct packages and GlobalPackageReference slipped past
the dependency check, the live-money scan missed the JSON, Dockerfile
and Unraid template forms, the secret scan missed short project
secrets but would have failed on the user's own ignored .env. The
scans now read only the files git would commit and cover those cases.
Also adds the fixed-seed generator for property tests, the REGISTRY
variables reserved for phase 9 in the CI, a git pre-commit hook in
place of the pre-commit tool (tasks hooks), the missing SPEC 4 facts
in API-NOTES (Fed/ECB/BLS times, USDC depth on OKX) and the full list
of today's indirect packages.
refs #3 refs #5 refs #6
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Empty but complete solution so every later phase starts from a working
toolchain: analyzers as errors, locked restore, format check, policy tests
that keep packages within ADR 0004 and keep live trading switched off,
a /health endpoint served from Docker, and a Gitea workflow with only
run steps. API facts of SPEC section 4 re-checked in docs/API-NOTES.md;
ADR 0006 records what changed (MSTest base packages, Ubuntu images,
OKX WebSocket on 443, Alpaca margin-only accounts).
refs #1 refs #2 refs #3 refs #4 refs #5
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user prefers C# and asked for no SDKs or extra components without
permission. Speed is not a factor at this trading frequency and Python's
advantages were exactly the SDKs now excluded, so ADR 0003 adopts the
appendix D variant with Blazor instead of React, in-house broker clients,
ADO.NET, .csv.gz research storage, a C# task runner instead of make and an
in-house Windows installer instead of Inno Setup.
ADR 0004 lists the only packages the user authorized (Microsoft.Data.Sqlite,
ProtectedData, WebView2, MSTest, Npgsql, ML.NET with LightGBM/FastTree,
ONNX Runtime, Tokenizers) and the ask-first rule for anything else.
ADR 0005 records why the bot is called Anubis (he weighs and protects) and
replaces the two-semicircle mark with the scale of the weighing of the
heart: SVG, PNG, ICO and Unraid icon, applied to all nine screens.
CLAUDE.md, SVILUPPO, STATE, README, ROADMAP and templates follow; the 84
Gitea issues were rewritten to match and verified against the source.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Load the full project specification renamed from the provisional "Giano"
to "Anubis" (ADR 0001), the nine reference screens with the new name,
the Unraid template draft pointing at the Gitea container registry, and
the working documents the development phases rely on: CLAUDE.md,
ROADMAP (linked to the 84 Gitea issues and 12 milestones), STATE,
development rules (SVILUPPO), ADR 0002 on Gitea issues, packages and
installer, and Gitea issue/PR templates. No application code yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>