sendMessage only checked hasBlockRelationship on the receiverId (DM)
branch; the chatId (group chat) branch skipped it entirely. A user
blocked by someone could route around the block by creating a
"group chat" whose only other member is the blocker and sending
through that chatId instead of receiverId.
createGroupChat now rejects chat creation if the creator has a block
relationship (either direction) with any proposed member.
Repro'd live against a disposable local Postgres+Redis: B blocks A,
A's direct message to B correctly 403s, but A's POST
/messages/chats + /messages{chatId} both succeeded (201) before this
fix. After the fix, chat creation 403s with "Cannot create a chat
with a blocked user"; unblocked group chat creation still works
(201). tsc --noEmit clean; full jest suite unchanged at the two
pre-existing failing files (messages.test.ts, messageService.test.ts,
identical 17/15 split confirmed via git stash against plain main).
`npm run dev` (nodemon --exec ts-node) crashed on startup with 8 TS2769
errors ("RateLimitRequestHandler not assignable to Application") every
time -- the documented quick-start path in README.md never actually
worked.
Root cause: this is an npm workspaces monorepo. express-rate-limit is
only hoisted to the monorepo root node_modules, not nested under
gitgram/backend/node_modules. Its own .d.ts resolves `express` by
walking up node_modules from wherever the package physically lives, so
it lands on the root's hoisted express@4 (and @types/express-serve-
static-core@4.19.8) instead of this package's own pinned express@5.1.0
/ @types/express@5.0.5 -- producing a real-looking but spurious cross-
package type mismatch on every `app.use(limiter)` call.
`npx tsc` / `tsc --noEmit` (the actual build and CI gate) do not hit
this at all and stay green -- confirmed before and after this change.
Set `transpileOnly: true` for the dev-loop ts-node process (the
standard fix for this exact class of ts-node + monorepo hoisting
false positive), so `npm run dev` boots instead of crashing on typings
that aren't a real bug in this project's own code.
Verified for real, not just via tsc:
- Spun up throwaway local postgres+redis containers, ran
`prisma migrate deploy` (4/4 migrations clean), booted the backend
with `npm run dev`'s actual command -- it now boots and serves
/health with a live DB connection; it previously hard-crashed before
a single request.
- Re-ran the full messaging + cross-tenant IDOR flow live end-to-end:
registered 3 real users, created a real group chat (A+B), both
members send/read successfully (201/200), and a non-member (C)
gets a real 403 on both GET /messages/chats/:chatId and
POST /messages with that chatId -- the IDOR fix from 487e4312 still
holds.
- Re-ran tests/messages.test.ts with and without this change (via git
stash): identical 17 failed / 15 passed both times, proving this is
the same pre-existing mocked-Prisma flakiness already documented in
plans/back-catalog.md, not a regression from this fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
explainTalentMatch's /explain endpoints 503'd because defaultRagAsk had no
real LLM client behind it. talentGraphLlmClient.ts tries Railyard Relay
(OpenAI-compatible, same tensorzero::model_name::anthropic:: pattern
bidstage-tenders already uses) first, falling back to a direct OpenRouter
call on any Relay failure or when Relay isn't configured.
Verified against real endpoints, not mocks: a real call through Relay
(claude-haiku-4-5-20251001) returned a grounded answer citing all 3 real
subgraph edges with zero hallucinated citations; a deliberately broken
Relay key correctly triggered the OpenRouter fallback path (confirmed via
a real 401 -> real fallback attempt -> real OpenRouter 402, no credits).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Audited gitgram-ai end to end by actually running the backend against a
live Postgres instance (not just reading/trusting the existing Jest
suite, which mocks Prisma and therefore never hits real FK constraints
or real route wiring).
Critical, feature-breaking bug: Message.senderId/receiverId and
MessageReaction.userId are foreign keys into Account, but every
messaging controller function wrote/read req.userId (a User id) into
those columns directly. Verified live: every message send (1:1 DM and
group chat) failed with a Prisma foreign-key-constraint violation, and
group messages additionally violated the (until now) required
non-null receiverId column since group messages have no single
receiver. Messaging has never worked through this code path. Fixed by
resolving the caller's Account id once per request (matching the
`prisma.account.findFirst({ where: { userId } })` pattern already used
elsewhere in the codebase, e.g. followRequestController) and using it
consistently for every Message/MessageReaction read, write, and
ownership check, while keeping ChatMember/UserBlock/socket rooms on the
real User id they were already consistent with. Added a Prisma
migration making Message.receiverId nullable for group-only messages.
Verified against a live Postgres test database: registered 3 users,
created a group chat, sent and read 1:1 and group messages, reacted,
and marked conversations read -- all succeed now.
Cross-tenant IDOR (same bug class found on Scribase today): GET
/api/v1/messages/chats/:chatId had no membership check at all -- any
authenticated user could read any group chat's messages by chatId
alone (every sibling function in the same file -- getMessage,
reportMessage, assertMessageAccess -- already enforced membership, this
one endpoint was the outlier). Also hardened sendMessage so posting
into a chatId you don't belong to is rejected. Verified live: a
non-member now gets 403 on both reads and writes to a chat they're not
in; an actual member still gets 200.
Unrelated build blocker found while standing up the backend to test
this: gitgram/backend/tsconfig.json set `ignoreDeprecations: "6.0"`,
which the pinned typescript@^5.9.3 doesn't recognize -- `tsc`/`ts-node`
hard-error with TS5103 before compiling a single file. Backend could
not build or boot in dev mode at all. Fixed to "5.0".
SEO gap: gitgram/web (Next.js 14 app dir) had no robots.txt and no
sitemap anywhere, static or dynamic -- zero crawl directives, no
sitemap reference for search engines. Added app/robots.ts (disallows
authenticated-only surfaces, allows public ones) and app/sitemap.ts
(static public marketing/discovery routes).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
DELETE /auth/delete-account threw "User not found" for every OAuth-only
account (any user with no password always failed `!user.password`), and
both the web and mobile clients called /v1/compliance/account and
/compliance/account, routes that never existed - so account deletion
was broken for 100% of users on every surface, not just OAuth.
Password accounts still require re-entering their password; OAuth-only
accounts delete using the authenticated session as proof instead. Web
and mobile now check whether the account has a password before showing
the field, and call the real /auth/delete-account route. Backend keeps
a DELETE /auth/account alias for any already-shipped client builds.
70/70 backend auth tests pass, including 3 new cases for the
OAuth-without-password path and the password-missing/wrong-password
cases.
The mobile app's settingsService.getNotificationPreferences/
updateNotificationPreferences call GET/PUT /api/v1/settings/notifications,
but the backend never registered any /settings/* routes — only
/api/v1/notifications/preferences existed. Every load silently 404'd
(swallowed by a catch that just logs) and every Save showed a "Failed
to save settings" alert; no notification preference was ever read back
or persisted for any user.
Mount the existing, already-working notificationPreferencesRoutes router
at /api/v1/settings/notifications as well, so the mobile client actually
reaches the real controller/Prisma logic instead of hitting a dead path.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
run-e2e.sh hardcoded the Compose v2 plugin (`docker compose`), which is not
installed on every machine/CI image (some ship only the standalone
`docker-compose` binary). Detect whichever is available so the docker-backed
e2e core-loop suite (real Postgres + Redis) is runnable everywhere.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
TypeScript 6.x makes moduleResolution=node10 (the legacy "node" resolver) a
hard error (TS5107), so `tsc --noEmit` exited 2 and the backend typecheck gate
was red. Add "ignoreDeprecations": "6.0" as TS itself prescribes; node10/
CommonJS resolution semantics are retained intentionally until a deliberate
resolver bump. Typecheck now exits 0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The default jest.config.js matched tests/e2e/*.e2e.test.ts, which are meant to
run only under jest.e2e.config.js against a real Postgres + Redis (npm run
test:e2e). Running them under the default (mocked) config made the parallel
suite non-deterministic: posts.e2e and snippets.e2e collided on shared DB state
and intermittently failed, while passing in isolation/serial.
Add testPathIgnorePatterns for tests/e2e so the default unit/integration run is
deterministic and needs no Docker datastores. The e2e suite still runs via its
own config. Default suite now green in parallel (57 suites / 1434 tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The recruiter placement/unlock path charges and settles through Creem
(creemService + signed creemWebhooks checkout.completed), but the
talentBillingService header, its settlement section comment, and the
controller's requestUnlock JSDoc still described a Polar adapter / Polar
order webhook. Correct those comments so the money path reads
unambiguously as Creem.
Verified before changing: @polar-sh/sdk and POLAR_WEBHOOK_SECRET remain in
active use by subscriptionService, storeController, and polarWebhooks for
subscriptions and the creator store, so they are intentionally left in place.
Comment-only change; no behavior change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The talent graph's `github_activity` evidence was written only by the GitHub
App webhook, so a new user's contribution-trajectory evidence stayed empty
until an owner configured a webhook and it fired — an ops dependency, not real
first-run data. Backfill it from the GitHub GraphQL contributionsCollection.
- githubContributionsService: pure `mapContributionsToActivities` turns real
commit/PR/review/issue contributions into RepositoryActivity rows, scoped to
the user's managed repositories (RepositoryManager), attributed to their
GitHub identity (actorLogin), deduped by a deterministic eventId that mirrors
the webhook delivery-id discipline. `harvestContributionEvidence` fetches via
octokit.graphql (no new dependency) and deduped-inserts with skipDuplicates.
Injectable db + graphql client so it is unit tested without Postgres/GitHub.
- githubController.syncProjects: run the backfill after project sync, non-fatal
(a harvest failure never breaks the primary sync), and report the counts.
This makes contribution evidence real and immediate after login with no
GitHub-App-webhook ops dependency and no fabricated seed data.
Tests: mapping (all 4 contribution types), managed-repo scoping, identity
attribution, deterministic-id dedupe, malformed/empty input, and the harvester
no-op / happy-path / graphql-failure paths. 20 new assertions across 10 tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The recruiter placement/unlock backend was end-to-end (routes,
talentBillingService, Creem checkout + signed webhook settlement) but had
no frontend, so recruiters literally could not pay. This closes the only
fully dead-ended money surface.
Web (gitgram/web):
- talentGraphService: add listUnlocks, requestUnlock (idempotent, optional
placement success-fee), and getCandidateContact mapping the existing
owner-gated backend routes, with typed placement/contact envelopes.
- utils/talentUnlock: pure gating helpers (deriveUnlockState, findPlacement,
interpretUnlockResult, formatFee) so the owner-only logic is unit tested
without rendering.
- talent-network page: owner-only "Unlock candidate" action on each match
card — opens the Creem checkout, reveals contact + full evidence once the
entitlement is paid (poll via "check payment" and on-return via query
params), with not-owner / pending / unlocked / already-unlocked states.
Backend (gitgram/backend):
- talentBillingController: point the checkout successUrl at the real
/talent-network page (with opportunity + unlocked params) instead of a
non-existent /talent/opportunities route, so payment return is not a 404.
Tests: unit-test the three service methods (idempotency header, success-fee
body, already-unlocked, 402-when-unpaid, id encoding) and the gating helper
(not-owner/unlockable/pending/unlocked, checkout-vs-reveal). 21 new tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The talent graph's recruiter monetization was designed (Opportunity → Pilot →
Proof funnel) but never wired to charge. This adds the placement/candidate-unlock
revenue engine on top of the existing talent graph, charged through Creem (a
merchant-of-record with Turkey-based-seller payout support — the portfolio
standard used by bidstage/CalDesk), not Polar/Stripe.
Revenue gate: a recruiter (opportunity owner) unlocks a matched candidate's
contact + full evidence. Gating mirrors the referral gate (owner-only, no
self-unlock, candidate must have real matching evidence). The fee is charged via
a Creem checkout whose custom_price is the exact computed fee, and settled by the
signed checkout.completed webhook, which records the entitlement + receipt.
- Fee math (pure, configurable): flat unlock access fee (TALENT_UNLOCK_FEE_CENTS,
default $49) or a percentage placement success fee (rate + min/max clamps) when
a placement amount is supplied.
- New TalentPlacement model + migration: entitlement/receipt with a compound
unique key (opportunity, candidate, recruiter) for charge idempotency and a
unique providerOrderId; provider defaults to creem.
- Creem adapter (createCreemCheckout) mirroring the reviewed bidstage lib/creem.ts,
and a signed Creem webhook route (HMAC-SHA256, timing-safe) that settles
placements. Amount is revalidated against the quoted fee before unlocking.
- Routes: POST/GET /talent-graph/opportunities/:id/unlocks and gated
GET .../candidates/:candidateId/contact (402 until paid).
- Errors.paymentRequired (402) helper.
- Tests: fee math, gating, charge idempotency, and the signed webhook settlement
path (75 tests). Secrets are env-only (CREEM_API_KEY, CREEM_WEBHOOK_SECRET,
CREEM_PRODUCT_ID) — none committed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Promote the talent graph's implicit structure to an explicit schema and add
the single guardrail that removes hallucinated structure: validate every
candidate edge against its edge type's domain/range and reject
type-incompatible ones.
- talentGraphOntology.ts: node types (developer, repo, org, skill,
contribution) and edge types (contributed_to, member_of, has_skill,
collaborated_with) with domain/range, symmetry and reflexivity rules.
validateEdge / validateEdges vet untrusted edges, require provenance
(source, observedAt, confidence) on every fact, dedup symmetric edges,
and return typed rejection reasons. Pure, no I/O.
- talentGraphEdges.ts: extraction step — turns getTalentGraph's view into
provenanced candidate edges, reusing canonicalizeSkill and the evidence
source union aligned with TalentEvidence.kind.
- tests: 15 unit tests covering domain/range rejection, provenance
validation, symmetric dedup, and extract-then-validate soundness.
Migration history reconciliation:
The autonomous-build migration history was corrupted — 20260109185732 dropped 13
tables still in the schema (Account, Advertisement, ScheduledPost, Chat,
Organization, Report, …) plus their enums/constraints without recreating them, so
a fresh `prisma migrate deploy` failed on the later repair/ALTER migrations
(ScheduledPost, Advertisement, AccountType, …). Squashed the 16 broken migrations
into a single clean 00000000000000_init generated from the current schema (93
tables, 19 enums, incl. MutedWord/MessageReaction/Post.contentLanguage). Verified:
fresh `migrate deploy` applies cleanly on an empty DB; the running DB was
re-baselined (resolve --applied init) and the backend boots healthy.
context7:
Switch the compose service from a broken raw node:20-alpine + read-only-volume +
npm-install setup to the package's upstream pnpm Dockerfile (build context =
context7 monorepo root), serving StreamableHTTP on container :8080 (host :3100);
point aitm-mcp's CONTEXT7_URL at :8080. Now Up/healthy.
Result: all 9 compose services healthy (gitgram backend+web, postgres, redis,
aitm-api/mcp/website, gitnexus, context7). E2e through containers still 201/200.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Replace the accidental from-empty migration (which duplicated all 16 prior
migrations and would break a fresh `migrate deploy`) with a proper incremental
migration adding only the new objects: Post.contentLanguage, the MessageReaction
and MutedWord tables, their indexes, and the MessageReaction foreign keys.
Verified the DDL against the live schema. NOTE: a fresh full `migrate deploy`
still fails on a PRE-EXISTING migration-history drift —
20260514174000_repair_scheduled_posts_schema references "ScheduledPost" which no
earlier migration creates. That predates this work; the running stack baselines
the schema (applied + migrations marked resolved) to sidestep it. Reconciling the
older migration history with schema.prisma is separate follow-up.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The whole stack now comes up healthy under one `docker compose` run (gitgram
backend+web + Postgres/Redis + aitm-api/mcp/website + gitnexus); only the
optional tools-profile context7 remains (pre-existing workspace-build issue).
gitgram/backend/Dockerfile: move to a glibc base (node:20-bookworm-slim). Alpine
musl lacks the glibc loader that @tensorflow/tfjs-node's native libtensorflow
needs (ERR_DLOPEN_FAILED), so NSFW detection now runs for real (no stubbing).
Full install WITH lifecycle scripts in the builder (downloads libtensorflow),
node_modules reused in the runtime stage so devDeps like swagger-ui-express are
present for the NODE_ENV=development compose run; openssl + ffmpeg + libgomp1
runtime libs; create writable /app/logs + /app/uploads for the non-root user.
prisma schema: binaryTargets -> debian/arm64 openssl-3 for the glibc runtime.
nsfwDetectionService: lazy-load tf + nsfwjs so a genuinely-unavailable native
lib degrades gracefully instead of crashing at import (still uses real tfjs).
gitgram/web/Dockerfile: npm ci -> npm install (workspace member's standalone
lockfile drifts after dependency additions).
Verified end-to-end through containers: register/login/post/feed 201/200,
web homepage + GitHub->Gitgram /caelum0x both 200.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
End-to-end smoke (real Postgres + Redis via docker-compose) caught a real bug
that mocked unit tests missed: createPost passed the request `type` straight to
prisma.post.create, so a client sending lowercase "text" (instead of the
uppercase PostType enum) threw PrismaClientValidationError 500. Now normalized
to the PostType enum (case-insensitive, falls back to POST).
Verified end-to-end: register→login→create post→comment→like→feed all succeed
against a live DB (3 posts / 1 comment / 1 user persisted).
Also adds prisma/migrations/20260629000000_marathon_additions (full schema incl.
MutedWord, Post.contentLanguage, MessageReaction) — resolves the deploy-time
migration that was infra-blocked from generating against the Neon pooler.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>