Browse this guideCapture project communication
Feature libraryUpdated 2026-11-17

Capture project communication

Provenience is designed to collect the communication and reference material surrounding a project without asking a model to guess where it belongs. SMS and MMS capture is available; the remaining capture paths are still planned.

One Site contact, arranged project context

Each Site will have one phone number and one email address for the business to use across its projects. A project's roster records its staff, customers, and collaborators. With a customer on a conversation, the customers' one shared active project decides, whoever else is on it; several shared active projects or none require an administrator's decision. With no customer, every registered party must be on the project: one shared active project files there, several go to the Review Queue with those shared projects as candidates, and none is company-level. A project is active until it is closed.

An unfamiliar phone number or email address is not assumed to be a customer. Its communication is captured, held from ordinary queries, and sent to the Review Queue for a business administrator to register or dismiss. A call about a closed project is also captured and queued; the administrator can file it on that closed project.

Projects may optionally receive their own phone number and email address; communication arriving on one is already arranged to that project. Calls go through an auto-attendant, are recorded on both sides, and are transcribed. Texts on the Site's group thread and email sent to its address will enter the same capture pipeline. A Twilio number in a native group thread receives each participant's messages with enough participant data to correlate the thread; outbound replies into that native thread are outside Phase 1.

Implementation progress

The SMS/MMS path now reaches the captured stage. Site ingress accepts an inbound Twilio callback, writes its original form body to the Site's Supabase Storage Deposit, and only then records it for later pipeline work. If Deposit storage is unavailable, ingress returns a temporary-unavailable response rather than recording an intake without its preserved original. If the later intake insert fails, it also returns a content-free 503 so Twilio can retry; that insert runs off the event loop. A new delivery and a duplicate both receive an empty <Response/> TwiML document with HTTP 200, which acknowledges the callback without sending a reply message. Ingress does not run model, roster, or other downstream processing while responding to the callback.

A separate worker turns each received row into a raw source. It records the original callback and, for MMS, retrieves every media item through an authenticated Twilio media request and preserves it as an attachment in the Deposit before marking the row captured. Extract is a later worker stage and is off by default; Capture can run without an Anthropic key. When Extract is enabled, it also records the text's participants and Contextualize files the resulting artifacts from the roster rule. An unfamiliar participant opens or joins an unknown-party Review Queue item; an unsettled customer conversation with artifacts, or a staff-and-vendor thread sharing several active projects, waits as project-undetermined. No message content is used to choose the project.

SMS and MMS capture takes a text from Twilio to the Deposit. Extract is off until an operator turns it on for a Site. Site-channel provisioning, access governance, and the remaining capture channels are still to come.

Connected sources

Two connector paths serve different kinds of source:

Both paths are read-only in Phase 1. Connector support will depend on what an integration agent can build against; there is no fixed vendor list.

Files

Files arriving through any channel are preserved and cataloged against their project. For SMS/MMS, Extract can read qualifying images and short PDFs together with the decoded message; every attachment also becomes a file record. Files outside the request limits are named but not sent to the model. Other file types, including spreadsheets and drawings, are preserved and cataloged without a claim that their contents were interpreted.

Current Site foundation

Every Site now has a private Deposit bucket in its own Supabase project, and a backend identity scoped to it. That identity can only add and read objects in the Deposit bucket — it cannot overwrite them, delete them, reach any other bucket, or bypass the database's row-level security, even though it authenticates through the same project as every other Site credential.

Deletion protection is structural rather than policy: the database refuses to update or delete a Deposit object or the bucket itself, and that refusal holds even against a credential otherwise powerful enough to bypass ordinary access rules. Nothing removes a preserved original today; an audited, tombstoned deletion path for retention is a later addition, not a bypass of this one.

Continue with Turn sources into governed knowledge.