Skip to main content
Meet (meet.tuturuuu.com, apps/meet) combines meeting planning and live collaboration. Public product explanations live on the main site; operational rooms, plans, and workspace meeting records are excluded from search indexing.

Choose the meeting surface

Before, during, and after a session

Before joining, verify the correct workspace/plan, participant access, and the configured meeting provider. During the call, recording, transcription, and assistant behavior follow the documented session controls and provider state. Afterward, review notes and proposed follow-ups before writing to Tasks or Calendar. A proposal is not an executed action, and a room URL is not broad workspace authority.

Transcription credit admission

Live transcription bills Gemini usage to the host’s personal workspace. Each upload window is admitted against AI credits before its transcript row is reserved, so a refusal never leaves a “[Missing audio segment]” gap. A failed credit check (RPC error or unreadable result) is retried at most three times with backoff, then returns a retryable 503 that stays within the client’s five-minute recovery window. A real refusal (credits used up, daily limit, model not on the plan, or another plan restriction) is a non-retryable 402 with a client-safe reason code. Meet then stops capture, finalizes the session, and shows the specific reason in English and Vietnamese instead of the generic partial-transcript alert. Notes generation keeps its own admission check. Regression controls: server/credit-admission.test.ts, server/chunks-credit-admission.test.ts and use-meeting-ai.test.ts under packages/meet-core/src/features/meeting-ai.

Public authentication callback

Meet’s native app renders the tokenless /verify-token and /vi/verify-token callback shell for exact GET and HEAD requests, including locale context, without consuming a verification token or requiring an account session. The explicit English locale prefix retains normal locale canonicalization. This lets a direct callback visit render the bilingual loading/logo UI and its hydration assets. An explicit token query parameter, including an empty value, and other methods continue through the shared verifier; token validation, safe next destinations, session cookies, login routing and protected routes retain their existing rules. The exception applies only to these exact callback routes, not API or workspace routes. Real-proxy controls live in apps/meet/src/proxy.public-callback.test.ts; SSR, hydration assets and both locales still require the emitted Worker CI gate.

Browser language selection

When a Meet page has no supported locale in its URL or locale cookie, browser language negotiation uses valid locale tags from Accept-Language in their quality order. Wildcard ranges and invalid locale tags are ignored before locale matching. If no valid preference remains, Meet uses its existing English default. Valid regional English and Vietnamese preferences continue to select their supported language. Zero-quality ranges remain excluded by negotiation. A supported URL locale takes precedence over the locale cookie, which takes precedence over the browser header. Public callbacks retain their existing locale handling; authentication, SSO destinations and safe next-path rules are unchanged. Real-proxy and Request-bridge regression controls live in apps/meet/src/proxy.public-callback.test.ts. Hosted original and restored Worker acceptance remains a separate gate.

Public login locale routing

For the exact locale-prefixed public login paths /vi/login and /en/login, Meet uses the configured English default for as-needed prefix handling. Explicit /vi/login continues to the login renderer with the Vietnamese locale header, even when the locale cookie or browser language prefers English. Explicit /en/login intentionally retains its redirect to unprefixed /login. Unprefixed /login retains the existing cookie and browser language selection; all other paths retain their existing locale handling. This path-scoped rule leaves the default authentication policy, public callback handling, token verification, session checks and branch ordering unchanged. Central SSO continues to return through Meet’s /verify-token callback, and the shared safe next-path guard continues to reject external destinations. The real-proxy/bridge regressions in apps/meet/src/proxy.public-callback.test.ts cover both login paths with absent, English and Vietnamese headers and conflicting locale cookies, preserving the encoded query, forwarded locale, exact callback continuation and zero fetches. Their fixed Start SSO response establishes only proxy/bridge continuation; shared login-page rendering and emitted original/restored Worker acceptance require separate runtime evidence. These controls do not establish the cause of a past hosted redirect.

Plan guest passwords and sessions

Scope: guests who join a Meet plan by name, with or without a password, on both the Meet satellite and the web-hosted plan page. Signed-in platform users are unaffected.
  • Storage. Guest passwords are stored with the shared server-only KDF in @tuturuuu/utils/crypto (scrypt, encoded as scrypt$N$r$p$salt$key with N=2^15, r=8, p=1, a 16-byte random salt and a 32-byte key). The salt lives inside the encoding, so password_salt is written as an empty string for new rows. Verification uses a constant-time comparison.
  • Credential flow. The guest submits the plaintext password over HTTPS to the guest login server action. The server verifies it and returns a signed guest session (@tuturuuu/utils/meet-guest-session) bound to that guest and plan, valid for 12 hours. Availability updates and guest poll actions present that session (guestSession); the server re-verifies the signature, plan, guest and expiry, and checks that the guest still belongs to the plan. Stored password hashes are never returned to the browser and are not accepted as a credential.
  • Default. Passwordless guests keep working: the empty password is hashed like any other value. The saved-name auto login on reload still applies only to passwordless guests; the password and the session are not persisted.
  • Plan access. Guest login, guest poll actions and guest availability writes (create and delete) follow the anonymous plan-access rule: they run getPlan(planId, { actorUserId: null }) before any session or database work, so a guest, including one holding a session issued earlier, only reaches public, non-workspace plans. Regression coverage: packages/apis/src/meet/actions/timeblocks.test.ts and apps/web/src/app/api/v1/meet/plans/[planId]/availability/route.test.ts.
  • Expiry. When a session expires the guest action is rejected and the guest signs in again from the plan page.
  • Attempt limits. Guest password attempts are limited per plan and guest name and per client IP (see shared-secret password budgets). While a budget is spent the form shows a localized retry message, even for the correct password. Verified logins refund one target attempt; source counters retain the charge. The source budget is 300 attempts per 15 minutes to tolerate shared office addresses. Creation reserves 10 attempts per plan and source network per 15 minutes, independent of the chosen name. Different networks do not share a plan-wide creation cap. IPv6 addresses share their /64; unknown IPs share a bucket. Missing plans do no hashing. Creation denials use distinct English/Vietnamese copy about attempts to join. The optional Durable Object password store is default-off. When assigned, each IPv4 /32, IPv6 /64 or shared unknown-source bucket, under the same pepper, may spend at most 5% of the password family’s daily command share across creation/login scopes, including rejected probes and successful refunds. This bounds one bucket, not a network allocation or one actor: rotating 20 /64s within one /48 can still drain the family. HIGH activation hold FAIR-R1-SYBIL-DRAIN remains open. Shared NATs and unknown-source callers can also exhaust their common bucket; P2 FAIR-R1-SHARED-SOURCE-LOCKOUT remains open. A repeated denied reservation is cached for at most one second. See the fairness activation holds. Regression coverage: apps/abuse-guard/src/fairness.test.ts and packages/utils/src/abuse-protection/__tests__/password-attempt-store.test.ts. The participant view and guest table are available only to server/admin clients; client roles cannot query either directly. Regression coverage lives in packages/apis/src/meet/actions/auth.test.ts, packages/utils/src/abuse-protection/__tests__/password-attempt-limits.test.ts, and apps/database/supabase/tests/security-meet-guest-grants.sql.
  • Table access. meet_together_guests keeps row level security with no client-role policies, and client-role table grants are revoked; only server code using the service role reads it.
  • Signing secret. Sessions are signed with the shared server token secret (TUTURUUU_APP_COORDINATION_SECRET, falling back to SUPABASE_SECRET_KEY) through a purpose-derived key, so they cannot be replayed as any other token type. Rotating that secret ends all active guest sessions.

Password hash migration

  1. Reset markers. verifyPassword accepts only the scrypt$ encoding. Any other non-empty stored hash, including the legacy unprefixed salted SHA-256 hex digest, returns resetRequired: true and never verifies or upgrades from a submitted password. Guest sign-in returns PASSWORD_RESET_REQUIRED, and the plan creator recovers the guest with Reset a guest password. The full policy is in shared-secret password reset.
  2. No backfill. Legacy rows cannot be converted without the plaintext, so the existing hash format is the durable reset marker; there is no schema change, migration step or bulk update.
  3. Parameter rehash. When a scrypt hash with older parameters verifies, the login request replaces it with a fresh scrypt hash using an update filtered on the previous password_hash value, so a concurrent upgrade is never overwritten. A failed upgrade is logged and the login still succeeds.
  4. In-flight clients. Browsers holding a page from before this change send the retired hash field, which the server now rejects. Those guests see a failed save or vote and sign in again after reloading.
Regression coverage: packages/utils/src/__tests__/crypto.test.ts, packages/utils/src/__tests__/meet-guest-session.test.ts, packages/apis/src/meet/actions/auth.test.ts, packages/apis/src/meet/actions/polls.test.ts, packages/apis/src/meet/actions/timeblocks.test.ts, apps/web/src/app/api/v1/meet/plans/[planId]/availability/route.test.ts, apps/meet/src/runtime/legacy-action-ports.test.ts and apps/database/supabase/tests/security-meet-guest-grants.sql.

Runtime boundaries

Meet shares behavior through packages/meet-core and packages/meet-ui. Realtime, signaling, provider connectivity, and hosted SFU/TURN are separate boundaries. A local Worker fixture does not establish hosted media-network success. The Meet runbook and linked call feature guide identify the owning deployment and verification path. Mobile workspace tools documents native availability and opt-in defaults. Rooms and plan share links stay noindex even when participants can join without browsing an entire workspace.

Native Worker verification

The native renderer uses TanStack Start SSR, Vite+ and the official Void SDK. Shared meeting behavior and the separate realtime Worker retain their existing ownership. Framework source or a successful build does not establish production cutover or hosted media acceptance. Preserve supplier-generated src/routeTree.gen.ts bytes, including generated casts. Meet and Calendar need owning-config formatter.includes set to the root Oxc ignorePatterns entries for **/routeTree.gen.ts and **/*.bridge.gen.ts. Preserve generated bytes and test authored files separately. Use bun run build:cloudflare in CI. Discover the emitted Worker configuration through exactly one of .wrangler/deploy/config.json or Void’s saved .void/preview-deploy-config.json; reject missing or simultaneous pointers. Void preserves the pointer bytes when moving it, so resolve configPath from the original .wrangler/deploy directory in either case. Do not assume an output directory or recreate a missing pointer. Preserve the existing Worker identity, targeted placement, bindings, secret names and Durable Object migration history. Qualify, seal and restore the complete output before publishing those exact bytes with the app-local Wrangler adapter. Void CLI deployment is not interchangeable: the released CLI rebuilds, prepares additional output and cannot preserve this targeted-placement path. Live qualification validates the materialized input and emitted config against Meet’s fixed resource policy and pinned supplier/parser profiles. It measures materializer and loader-selection facts twice, rejects typed-source, input or profile drift, and admits only explicit default representations. Empty queue producer/consumer lists cannot authorize a queue; existing bindings, migration history and placement remain exact. Inspect the bounded field-rejection reason when qualification fails; diagnostic absence does not approve a value. The schema-2 artifact seal retains the original input path/hash, platform and qualification profile alongside exact output hashes. Restored discovery requires the seal, external expected SHA and complete file hashes before policy checks. It compares historical metadata without opening or rebasing the original input path and never falls back after a malformed seal. Raw emitted config bytes stay unchanged. This relationship does not prove historical retained-state provenance. Regression coverage is in scripts/ci/verify-meet-start-output.test.js. Both the original prepack dry-run and production restored dry-run use scripts/ci/meet-dry-run-scratch.mjs from apps/meet, passing "$SOURCE_SHA". The wrapper derives the emitted config through the existing artifact verifier, requires its manifest and canonical output, and captures the strict directory seal and absence of the config parent’s .wrangler and .wrangler/tmp before running the same app-owned bun x --no-install wrangler deploy --dry-run -c command with the discovered config and inherited environment. Wrangler can leave empty .wrangler/tmp ancestors after removing its deploy child. After successful child close, the wrapper accepts no residue or exactly that empty two-directory layout. It verifies file hashes and the complete prior layout before using nonrecursive rmdir on the leaf and then parent, and rechecks the manifest and strict directory seal afterward. Preexisting, partial, nonempty, linked, special or unexpected entries fail closed; command failure preserves residue without cleanup or retry. The file manifest alone cannot reject extra empty directories. Keep the hidden-entry allowlist, full tar contents and launcher’s original-artifact seal unchanged. The credential-free restored gate does not need another dry-run or rebuild. On rejection, preserve the failed artifact/evidence for bounded inspection; do not recursively delete scratch or normalize the archive. Syntax and source review do not qualify this repair: require exact-commit dry-run, pack/restore, and original/restored launcher acceptance in CI before claiming producer repair. After qualification, collect native-emitted-config-meet.json even on failure. Keep the actual source HEAD separate from the event SHA. Retain pointer and config hashes and a redacted key-shape inventory, never raw configs, resource identifiers, or secret values. Unavailable or overflow diagnostics remain failures. The receipt provides no acceptance, allowlist approval, or rollback proof. Inspect the actual receipt before qualifying unexpected queues or configPath fields. Preserve the strict resource, saved-pointer, native-relative-base, restore, and deployment gates. Diagnostic receipts use schema 2 with wireEncoding: "scoped-layouts-v1" when inventory or defaultProbe is present. Only those fields use shared ordered layouts: [0, layoutIndex, ...values] represents an object and [1, ...items] represents an array. Scalars and full hashes remain unchanged; optional absence, null, false, property order and row order are preserved. Top-level status and provenance remain readable. Use decodeReceipt from scripts/ci/native-receipt-layout.mjs to restore the original logical schema 1; it also accepts legacy schema-1 receipts. It rejects unsupported encodings, malformed layouts/tags, duplicate or prototype-related layout keys, excessive depth and expanded size. Decoding never qualifies acceptance or supplies missing status/evidence. inventoryHash still commits to the original inventory JSON. The complete wire receipt, including its final LF, must fit 16,384 UTF-8 bytes. Overflow retains the existing schema-1 unavailable summary and failing exit code, without partial diagnostics. Collection and the live default probe stay logical objects. Regression coverage lives in native-receipt-layout.test.mjs and native-emitted-config-receipt.test.mjs; actual emitted-payload fit requires CI on the new commit, separately from historical and synthetic fixture coverage. Request adapters preserve satellite sessions, locale rewrites, response cookies and central API ownership. Native plan mutations use POST server functions with CSRF checks and the canonical authorization suppliers. Intentional credentialed guest actions remain scoped to the existing plan policy. Confirmed mutations invalidate the native loader; denied or uncertain results do not report success. The owning workflow must verify emitted server/client boundaries, bilingual SSR, SSO redirects, logout cookies, API forwarding, static assets and Live admission. The artifact-dependent suite is scripts/ci/meet-cloudflare-render.runtime.mjs; it runs after building and again against the restored artifact, outside generic source-only test discovery. The credential-free restored-artifact job depends on validate and downloads its exact-SHA meet-worker archive into a fresh immutable-source checkout with frozen dependencies. Unlike the original post-build gate, it restores the uploaded bytes without rebuilding, verifies the seal and discovers the restored config, then runs scripts/ci/meet-native-launcher.mjs and the finite SDK checks under the existing three-minute deadline. It retains finite graph evidence even on failure and rechecks the manifest after acceptance. This job has only contents: read, no production environment or account credentials, and performs no deployment. The production job retains its own restored-runtime defense. Workflow wiring is not an executed CI result: client-boundary source review does not establish hydration or token completion. Provider, staging, canonical-route and recoverable rollback acceptance remain separate requirements. Signed-in workspace, room and plan flows, all native mutation transports and browser hydration need independent acceptance. Preserve an observed hosted configuration and recoverable prior Worker version before traffic changes. Local workerd probes do not qualify hosted SFU/TURN, devices, recording or assistant stop behavior. The consent log is private server data. Neither authenticated nor anonymous clients can insert, update, or delete evidence; the server appends decisions and orders them by created_at DESC, id DESC. Only the service role can erase rows. A withdrawal cannot be removed by a client to revive an earlier allow (F7). A BEFORE UPDATE trigger also refuses privileged evidence edits; only nested server/Auth account-deletion actions can clear user_id, retaining all other fields. A random UUID participant_ref is minted on first consent per session/participant, reused for later decisions in that pair, and stored only on consent rows. It is neither a user-ID digest nor an email/name hash. Deleting an account sets the consent user_id to NULL while retaining that random pseudonym and the evidence. Clients cannot delete Auth accounts or trigger this anonymization. Auth account deletion clears consent identity immediately, even though the legacy Auth trigger retains the public profile. Later service profile erasure removes its personal workspace after Auth deletion; consent evidence survives both steps. These decisions apply by default to Meet and Parley session consent, including live and recording sources, with no client-role exception. Meeting deletion and account deletion are distinct from an explicitly authorized evidence erasure. They do not expand transcript purge into evidence deletion. Guardian/legal retention policy remains an independent product/legal decision. Meeting/session deletion is refused while consent rows exist (ON DELETE RESTRICT). Only a trusted service erasure flow explicitly deletes consent rows before deleting the session or meeting. Transcript retention purges text in place and retains the session and consent evidence; it needs no consent erasure. This schema does not enable audio storage. Meet and Parley retain host-only visibility for existing sessions and callers that omit visibility. Workspace settings default new application-created sessions to participants, explicit per-person opt-in, an ask prompt, zero audio retention, and 365 days of text retention. Retention is bounded to 1–365 days or NULL (forever); audio retention is enforced at zero until audio storage is implemented. Notice versions are capped at 80 characters. Workspace settings enforce required account MFA in addition to membership/admin RLS. Applying workspace overrides belongs to the server integration; adding settings does not bypass participant consent. The M1-B migration explicitly replaces private.purge_meet_transcripts_batch: purge_after <= now() selects expired sessions in both dry-run and delete; NULL retains a session forever. Existing deadlines are backfilled from coalesce(ended_at, created_at) + 365 days. New sessions default to 365 days from creation. Purging clears both original and edited notes, and transcript chunks; legacy recording transcripts retain their separate 365-day policy. The migration and rollback close the purge fence and withdraw delete approval. A database owner must approve fresh dry-run counts before enabling deletion. The existing batch, daily, terminal-state and stop-fence limits remain in force. Regression coverage: meet-ai-consent-schema.sql checks client denials and withdrawal ordering; meet-ai-consent-retention.sql checks creator/admin deletion refusal, server erasure ordering, random pseudonym reuse and account anonymization; required-account-mfa.sql checks MFA policy. meet-transcript-retention-purge.sql checks custom deadlines, forever retention, edited-note deletion, ACLs and bounded restart behavior.