> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tuturuuu.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Parley

> Operate the private scenario training application and shared Meet runtime.

Parley runs at `parley.tuturuuu.com`. It uses Tuturuuu app-session authentication
and the shared `@tuturuuu/meet-core` runtime. Verified `@tuturuuu.com` accounts
have access; external accounts require an enabled entry managed in Infrastructure
under **Platform administration → Parley**. Authorization checks the current
confirmed account identity. Failed lookups deny access.

## Private content

Infrastructure administrators manage scenario briefs, private AI instructions,
roles, debrief rubrics, publication state and reference uploads. Reference bytes
are stored in `private.parley_references`, accessible only through authorized
server operations. Source documents are never seeded or committed. Import real
content through Infrastructure separately in each intended environment.

Starting a session atomically creates a Meet room and an immutable scenario
snapshot. Later edits do not change a running session's instructions. The
facilitator can save observations, decisions and debrief notes; participants do
not receive these private notes or raw source documents. Discuss consent before
recording or collecting research observations. Notes describe observable actions,
not validated personality or truthfulness assessments.

Mira uses the stored scenario instructions for shared chat and live voice.
Parley's lightweight avatar follows the local playback audio envelope. It does
not analyze participant faces or microphones, and it is not phoneme-accurate
viseme synthesis or a full 3D avatar.

## Development and deployment

1. Apply the additive Parley migration locally and regenerate database types.
   Do not reset a shared development database to test this app.
2. Run `bun --cwd apps/parley dev` with the same Supabase, app-session signing,
   realtime and provider environment contracts as Meet. The app uses port 7834
   and `parley.tuturuuu.localhost`.
3. Validate the shared runtime and build Meet, Parley and Infrastructure through
   `ttr resources run`. The extraction affects all three consumers.
4. Sync the reviewed release to production through the authorized release
   workflow. Verify that the production database migration job applies all
   Parley migrations. Meet and Parley Worker deployments wait for the database
   readiness marker; do not run a manual production schema push. Publish the
   central auth registry and Infrastructure management surface before testing.
5. Build Parley with `NEXT_PUBLIC_MEETING_APP=parley`. Its web Worker and cache
   bucket are separate from Meet. A private `MEETING_AI_PROVIDER` service binding
   calls Meet’s named provider entrypoint; `MEET_LIVE` references Meet’s Durable
   Object namespace. The Gemini credential stays in Meet. Deploy Meet first,
   then Parley; CI waits for the shared runtime readiness endpoint. The common
   realtime signing secret is provisioned from the protected GitHub environment.
6. Verify denied, internal-domain and allowlisted sign-in; revoke an external
   account and recheck access. Verify private reference upload/retrieval, scenario
   persistence after reload, room admission, two-participant audio, Mira response,
   and audio-driven avatar motion in each environment.

Existing live sockets use short-lived credentials. Revocation prevents new
requests and tokens; it does not retroactively disconnect an already-open call.
Close the session when immediate termination is needed. Do not describe a build
or synthetic test as production or physical multi-device verification.

## AI Hub and billing

Meet and Parley use the same workspace credit system. Live voice reserves credits before connecting to the provider, settles cumulative modality usage, and refunds the unused reservation. App and room identifiers accompany the reservation and final ledger entry, without scenario text or private documents.

Transcription and meeting notes reserve an AI Hub run before generation. They use the purchased output-token limit, settle measured provider cost with the plan markup, and reuse the same run for settlement retries. Missing provider accounting fails closed and leaves the reservation visible for reconciliation. These runs appear in AI Hub observability; chat also reserves and settles an AI Hub run, while live deductions appear in the combined credit ledger. Chat settlement consumes its own hold, including near the quota limit; optional public search is disabled when its additional reservation would exceed the available credits. The usage and credits links open the payer's personal workspace.

Apply the research migration plus the meeting attribution and observability migrations before deploying either app's new billing code. A Gemini key marked sensitive in Vercel cannot be exported. Parley uses the private provider binding instead of duplicating that key. Native local development may configure its own development key; do not bind local testing to the production provider.

Parley disables public Google Search during scenario roleplay. The scenario configuration is separate from the participant request, so reference text never becomes an explicit public search query.