Health: public gets only ok/error; details need X-Internal-Key. Docs/schema: admin session only.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Found on the live server (3 MB JSON to /api/auth/login answered 500). Django raises
RequestDataTooBig for bodies over DATA_UPLOAD_MAX_MEMORY_SIZE and would answer 400, but our DRF
exception handler turned every unknown exception into 500 + logger.exception, so each large request
also wrote a full ERROR traceback. Other SuspiciousOperation subclasses now get Django's 400 as well.
Tested: 2 tests (the 413 one fails without the change). Verified on the server after deploy.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The owner decided (2026-10-07) that we do not connect to anything other teams run on the box, Redis
included. CACHE_URL was emptied on the server and our keys were removed from the shared Redis.
- settings: drop KEY_PREFIX / CACHE_KEY_PREFIX (it only existed because the Redis was shared); comments
say CACHE_URL is intentionally empty and ResilientRedisCache is ready for our own Redis.
- config/cache.py docstring and tests updated accordingly (11 tests, still run against a local redis-server).
- CLAUDE.md: new server rule 8 (touch only our own files and service), status block, historical line
that said "db1 is free" now says not to connect. README, operations, handbook, security doc (item 9
reopened as an open decision) say the same.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Since 2026-10-07 the cache is the shared Redis on the box. Django's RedisCache raises on any network
error and DRF reads the cache for every throttled request, so a Redis outage meant 500 on almost every
route, /admin/login/ and OTP request included (all seven drill tests fail on the stock backend).
config.cache.ResilientRedisCache turns Redis errors into logged (rate-limited ERROR) misses:
get->default, set/delete->ignored, add->True, incr->ValueError. Cache-based caps (DRF throttles, hourly
OTP window, admin-login limiter, payment reconcile lock) are NOT enforced during an outage (fail-open);
DB-backed caps stay. /api/health returns 503 meanwhile. KEY_PREFIX namespaces our keys on the shared
Redis; 1s connect/read timeouts stop a hung Redis from holding requests for gunicorn's 120s.
Tested: 11 tests, incl. a real local redis-server killed mid-test and restarted without restarting the
app. Not tested yet on the server at the time of this commit.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- CLAUDE.md: status block for 2026-10-07 (what changed on the server, how it was checked, what was not).
- docs/security-before-production.md: signing key, shared cache and unit DEBUG items moved to the guard table.
- requirements.txt: ranges replaced by the versions running on the server (pip freeze, 2026-10-07).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
From the server the public URL goes through ArvanCloud and returned 504/000 (2026-10-07), so the
final health step failed even for a healthy service. It now asks the gunicorn socket (path read from
the unit) with our Host header first. DEPLOY_BUNDLE=<file> fetches from a git bundle instead of Gitea,
only when set explicitly; without it an unreachable Gitea still stops the deploy.
Tested: probe_health against a throwaway unix-socket server and a dead URL (bash -n clean).
Not tested: the whole script end to end, until the next deploy.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- content: POST nodes/{id}/quiz/complete and lessons/{id}/complete refuse a locked island
(403 node_locked, same rule as /api/home) and a user who has not answered every question
(409 answers_missing). Before, only a user with zero hearts and no answers was refused, so
anyone else finished any island, locked or not, and finishing a locked island's lessons
unlocked it. A node or lesson with no questions still passes when unlocked, with no XP.
The client can no longer understate totalQuestions to buy stars.
- throttle_scope on 15 routes that had none (question bank without login, intro, hearts,
notifications, tickets, plans, subscription status, orders, offers, affiliate, assistant
history and feedback, logout).
- admin: OTPCode and Session can only be added/changed by a superuser.
- deploy.sh: integrity-checked DB backup before migrate (keeps the last 2); no migrate without it.
- docs and CLAUDE.md updated to say what is verified and what is not.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
accounts.0008 drops the columns and refuses to run while guest rows exist.
Docs describe the current state; tests that only proved guest behaviour are gone.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
POST /api/auth/session is gone (404). Existing is_guest rows keep refresh and
upgrade-by-phone until the demo/guest rows are purged; the code that serves
them is listed in docs/security-before-production.md §10. Tests that used the
endpoint as a fixture now use accounts/testing.guest_session.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- accounts admin: phone/is_guest read-only for non-superusers, device_id filter lookups refused,
auth.User/Group admin writable only by superusers
- campaigns: delivery log needs view permission
- phone/OTP/birth date: any Unicode digit normalised or rejected with 400; unknown gradeId is 400
- shared demo guest is never turned into a person's account by logging in
- /api/home with unknown gradeId is no longer cached (LocMemCache flood reset throttle counters)
- /api/progress/nodes: one query, huge unit is 400
- BazaarCredential token filter lookups refused
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- Bazaar verify: a rejected token is not stored on the order or the failed transaction, so the
same purchase can still pay the right order; a refunded order refuses a new token (409).
- OrderAdmin: every field read-only, no add/delete; Transaction admin read-only.
- grant_pack uses the order's snapshot when the plan was deleted, logs ERROR when nothing is granted.
- Pay link: authority is stored only if the order amount did not change during PaymentRequest.
- PlanAdmin accepts only IRT (the app always renders toman).
- reconcile_pending_order logs unexpected exceptions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Admin: refund/cancel/mark-resolved/Bazaar-link actions need change permission; question and
curriculum import need add/delete; content browser needs view_question; guest device ids are
shown to superusers only.
- Accounts: registering a guest's phone revokes the guest's earlier sessions; self-deletion keeps
Redemption, UserOffer and the phone's OTP rows; the public demo guest cannot be deleted
(DEMO_GUEST_DEVICE_IDS); display-name contact filter covers any separator and social handles.
- Payment: callback without Authority is 400; a second Bazaar purchase on a paid order is 409;
orders snapshot credits/hearts/days at creation (payment.0011).
- Assistant: a broken diagram no longer makes a streamed answer free; a late off-topic marker
refuses the answer (not stored, not cached); a disconnect saves the partial answer; only
KNOWN_COLLECTIONS are searched.
- Infra: secret-key file is created atomically (workers no longer split keys); no default
INTERNAL_SERVICE_KEY; Kavenegar key and Bazaar token kept out of error logs; Swagger/ReDoc
pinned; schema throttled; docker compose bound to 127.0.0.1; non-ASCII digit ids are 400;
ticket text capped at 5000.
- docs/security-before-production.md: remaining open items 24-31 and the guard tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- /api/payment/orders/{id}/pay requires the short-lived signature that paymentUrl carries;
without it (or with another order's) it is the same 404 as an unknown order
- Session.refresh_token stores the SHA-256 of the token; migration accounts.0007 hashes the
existing rows so tokens clients already hold keep working
- /api/leaderboard requires a token (guests have one); names of children were scrapable
- RAG mock only for `manage.py test`, not any command with an argument named "test"
- tools/loadtest: user factory, load generator, fake GapGPT+Qdrant, with method notes
- security-before-production.md lists only open items; closed ones become a guards table
with the tests that hold them. Comments describe current behaviour, not history
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first assistant round after a restart waited ~6 s before reaching the model on the
live server (later rounds ~1 s); locally a cold 10-ask burst took 7.5 s vs 6.1 s with
this hook (simulated model, 5.2 s). Workers are recycled by max_requests, so the delay
recurred. A failed warm-up is logged and does not stop the worker.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- gunicorn gthread 8 -> 32 threads per worker: each assistant stream holds a thread for
the whole answer; with 3x8, about 20 concurrent asks already queued ordinary page loads
for seconds (up to 23 s at 50 asks locally, 7 s at 20 asks on the live server). With 3x32,
75 concurrent asks kept page loads under 110 ms (laptop, simulated model and Qdrant)
- deploy.sh: new step closes the project dir to other accounts and sets the database and
backups to 0600/0700 on every release; user `nobody` on the shared server could read
db.sqlite3 and every backup
- requirements: python-dotenv 1.2.2, urllib3>=2.8.0 (known CVEs in the installed versions),
gunicorn pinned to the server's version (was only installed by hand)
- security-before-production.md: items 19-23 (file permissions, DEBUG=1 still in the systemd
unit, dependencies, frontend headers, guest AI cap per shared IP) and the reproduced
X-Forwarded-For spoof through the origin IP
- model-prices.md: 0.183 cents per answer from 50 answers in the live load test
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- seed_content no longer undoes admin work: a SeedRecord ledger (content/seed_sync.py)
stops deleted/renamed/re-worded content from being recreated and keeps admin-edited
lesson text; new rows in the seed files still arrive (migration content.0010)
- plan list cache is invalidated on Plan save (receiver used a stale 'v1' key)
- token refresh per-IP cap 30/hour -> 600/hour (TOKEN_REFRESH_IP_RATE)
- a failed gateway result only changes a pending order (forged NOK no longer
overwrites refunded/canceled)
- tutor gets the correct answer only after the student has answered
- support reply notification moved to a TicketMessage post_save signal
- global daily cap on paid AI questions (AI_GLOBAL_DAILY_MAX, default 3000)
- Bazaar verify uses only the order's SKU; bazaar orders require a SKU
- /api/health warns on the committed signing key; deploy.sh saves local commits;
400 for bad ?unit= / ?period=; account delete keeps the referrer's Conversion
- earlier today: no stub answers outside DEBUG, ?grade= removed, logged fallbacks
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- /api/ai/ask: the stub answer is reachable only under DEBUG or
RAG_FORCE_MOCK. Without RAG keys on the server it used to return 200 with
"this is a stub, set OPENAI_API_KEY" and spend the student's quota; now 503.
- /api/home: the `grade` alias is gone and sending it is a 400 instead of a
silent fallback to the user's own grade. No client sent it.
- activate_subscription logs ERROR when the paid order's plan no longer exists
(it silently became 30 days).
- resolve_grade_title logs instead of `except: pass`; curriculum_io no longer
reads an `intro_text` key no file contains.
- CLAUDE.md: "No silent fallbacks" rules for future agents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- technical-handbook: question packs instead of subscriptions, real
hearts with the new wait screen (screenshot), math-only tutor and what
is charged (new 9-7), admin cancel/refund, Bazaar planId, account
deletion, 59 endpoints, 779 tests, 14 apps, status and open decisions.
- Diagrams redrawn: use-case (buy a pack, refunds), data model
(QuestionCredit, HeartsPass, Profile.birth_date), pack payment sequence.
- The PDF is now built from the Markdown by docs/build_handbook.py
(pandoc + Chrome, font in docs/assets); the hand-made HTML it used to
come from was not in the repo and had been lost.
- architecture, operations, security-before-production (item 13 closed),
ai-cost-and-guardrails and README follow the same changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was no way to take an order back: staff could only edit the
status by hand, and turning 'paid' into 'failed' left the credits,
heart days or subscription days the order had given. Two OrderAdmin
actions now do it:
- cancel a pending order (a payment that does arrive later still counts);
- refund a paid order, after a confirmation page: takes back the pack's
unspent credits (never below zero), its heart days, or the plan's
subscription days, and undoes the referrer's commission.
The money itself is returned in the Zarinpal / Bazaar panel; the action
says so. Status is read-only in the order form. Steps in
docs/operations.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The zero-hearts check moved inside the answer transaction: two wrong
answers sent at once with one heart left were both accepted (reproduced
with two threads).
- Finishing a lesson or an island without hearts and without having
answered its questions is now 403 no_hearts; before, quiz/complete let
a student move on island after island with zero hearts. Losing the last
heart on the last question still finishes the island.
- A heart that came back while waiting is now saved before answering, so
the reply no longer says 0 hearts left (the app then showed the
no-hearts page to someone who had a heart).
- The 403 also carries fullRefillAt for the waiting screen.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What GapGPT actually billed per model, with dates and evidence, so a
later price rise can be compared with what it was. gpt-4.1-mini and
gpt-4o-mini were re-checked from the 852 recorded calls in
docs/benchmarks/2026-09-26-tutor-model: every call equals the official
OpenAI per-token price or less (input cache), none more.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Plan.kind (pack/subscription) collided with the lesson kind
(quiz/video); drf-spectacular named one KindEd8Enum and warned.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The claim that the user-row lock plus SQLite IMMEDIATE stops one credit
being spent twice was untested. Two threads on separate connections now
ask at the same time: one answer, one credit spent; the same
Idempotency-Key arriving twice at once also costs one credit. Removing
transaction_mode IMMEDIATE turns both red (500, database is locked).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bazaar orders are only created through POST /api/payment/orders, which
took no planId, so a pack order had plan_id=None: Bazaar verified the
purchase, the order became paid, and grant_pack never ran. The endpoint
now takes planId (rejected when unknown) and snapshots the plan title.
The verify response also returned the pre-payment status ('pending')
because apply_order_payment_result updates its own copy of the order.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
