CL-7107: re-pin Interchange vendor to 692c3106

ARCHITECTURE.md

@@ -30,7 +30,8 @@ this repo:
 - **Tenancy** — the tenant hierarchy, membership, principals, roles, and
   grants (`@intx/db`, `@intx/hub-api`).
 - **Credentials** — resolution and storage.
-- **Agents** — launch and session orchestration (`@intx/agent`,
+- **Agents** — Interchange owns launch (`prepareProvisionedDeployment`),
+  the run principal, and session primitives (`@intx/agent`,
   `@intx/hub-agent`, `@intx/hub-sessions`).
 - **Inference** — LLM calls (`@intx/inference`).
 - **Workflow runtime** — definitions, runs, and the workflow host
@@ -44,6 +45,36 @@ for the authoritative list of what Interchange already provides versus
 what is a genuine upstream gap workbench has had to work around
 product-side.
 
+## Agent launch
+
+Workbench is a client of Interchange. Launching an agent is Interchange
+`prepareProvisionedDeployment` onto that instance's own `kind:workflow`
+asset (default ref) in the hub git store. Interchange mints the run
+row, the allocation, and — on first trigger — the run principal. Chat
+records the returned run id, deployment id, and mail address. It does
+not pre-mint an anchor run, own an `agent_session`, or render a
+per-run `sourceRef`. `@corbits/folded-runs` and
+`@corbits/folded-run-one-shot` are gone. A one-shot wait-loop for Myra
+drafting lives in `@corbits/agent-directory` as `runOneShotPrompt` and
+still calls `prepareProvisionedDeployment`.
+
+Launch is asynchronous. The sidecar reconciler's ready hook and the
+`workflowDispatchService` queue (wired in `apps/hub`) deliver a
+message sent before the instance is ready; the UI shows a starting
+state until then. Presence while starting comes from allocation and
+deployment status, not a run principal. Wake of a reaped or lost
+instance is a fresh `prepareProvisionedDeployment` onto the same
+asset; chat updates the recorded ids.
+
+A turn is one native `sendUserMessage` per addressed agent, carrying
+Message-ID, In-Reply-To, and References. Agent replies reach humans
+through the hub `persistMail` wrapper into `@corbits/mailbox`. Turns
+do not use a correlationId. Human attachments are read from the
+mailbox frame, not Interchange `session_mail`. Inference sources ride
+as catalog offering ids (`sourceOfferingIds` /
+`defaultSourceOfferingId`). Catalog instantiate does not 409
+credential-binding workflows.
+
 ## Tenancy model
 
 A **bench** (the shared team scope) and a **workbench** (a single

PRODUCT.md

@@ -46,8 +46,9 @@ column at a time:
   things to join. Plus mints an empty channel. Nobody is auto-hosted.
 - **Agents are principals, not templates.** Opening Sales opens Sales —
   the one 1:1 tenant with that agent. The same agent can sit in its DM
-  and in many channels. Product reopens or invites; it does not clone
-  the definition or mint a sibling instance per room. Myra is the
+  and in many channels. Inviting an agent provisions a native
+  Interchange run of that definition; product reopens or invites, it
+  does not clone a sibling instance per room. Myra is the
   first-run guide in her DM, not a special home slot and not a
   parallel home route. When she creates a specialist (`create_agent`),
   the default is to mint or reopen that specialist's own DM

apps/hub/package.json

@@ -28,8 +28,6 @@
     "@corbits/docker-provisioner": "workspace:*",
     "@corbits/e2b-sandbox-sidecar": "workspace:*",
     "@corbits/evals": "workspace:*",
-    "@corbits/folded-run-one-shot": "workspace:*",
-    "@corbits/folded-runs": "workspace:*",
     "@corbits/inbox": "workspace:*",
     "@corbits/inference-catalog": "workspace:*",
     "@corbits/insights": "workspace:*",
@@ -43,7 +41,6 @@
     "@corbits/process-provisioner": "workspace:*",
     "@corbits/run-key-history": "workspace:*",
     "@corbits/run-scope": "workspace:*",
-    "@corbits/sidecar-placement": "workspace:*",
     "@corbits/skills": "workspace:*",
     "@corbits/slack-tag": "workspace:*",
     "@corbits/tool-registry-publish": "workspace:*",

apps/hub/src/config.ts

@@ -272,7 +272,6 @@ function parsePositiveMsEnv(
 
 const SEED_MODEL_PROVIDER = "anthropic";
 const SEED_MODEL = "claude-sonnet-5";
-const SEED_MODEL_BASE_URL = "https://api.anthropic.com";
 
 // Matches boot-time seeding's own defaults (`system-seed.ts`) exactly,
 // so a zero-.env-edit local checkout that seeds its admin account
@@ -282,10 +281,14 @@ const DEFAULT_PLANT_ADMIN_EMAIL = "alice@example.com";
 const DEFAULT_PLANT_ADMIN_PASSWORD = "password123";
 const DEFAULT_PLANT_ORG_SLUG = "workbench";
 
+/** The root tenant's boot-time seed model: a provider/model pair
+ * `seedTenant` names in every deployed definition, plus the real (or
+ * placeholder) key `seedCatalog` plants a launchable credential with. No
+ * `baseURL` — a workflow deploy resolves inference from the tenant's
+ * catalog offerings, never a bare source tuple (CL-7461). */
 export type ModelSource = {
   readonly provider: string;
   readonly model: string;
-  readonly baseURL: string;
   readonly apiKey: string;
 };
 
@@ -624,7 +627,6 @@ function seedModelFrom(parsed: ParsedHubEnv): ModelSource | undefined {
   return {
     provider: SEED_MODEL_PROVIDER,
     model: SEED_MODEL,
-    baseURL: SEED_MODEL_BASE_URL,
     apiKey,
   };
 }

apps/hub/src/crypto-provider-cache-wiring.test.ts

@@ -74,6 +74,6 @@ describe("hub crypto-provider cache wiring", () => {
     expect(firstCall(source, "launchWebhookTrigger")).toContain(
       `cryptoProviderCache: ${assigned}`,
     );
-    expect(firstCall(source, "runOneShotFoldedPrompt")).toContain(assigned);
+    expect(firstCall(source, "runOneShotPrompt")).toContain(assigned);
   });
 });

apps/hub/src/hub-error-handler.test.ts

@@ -6,7 +6,7 @@
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
 import { Hono } from "hono";
 import { configureSync, resetSync } from "@intx/log";
-import { InferenceResolutionError } from "@corbits/folded-runs";
+import { InferenceResolutionError } from "@corbits/chat";
 
 import { hubErrorHandler } from "./hub-error-handler";
 

apps/hub/src/hub-error-handler.ts

@@ -16,7 +16,7 @@ import { makeErrorEnvelope } from "@corbits/error-sink";
  * Duck-typed rather than an `instanceof` allowlist: any error carrying a
  * string `guidance` alongside its `message` is a named, consumer-facing
  * failure the thrower already wrote for a human to read — see
- * `@corbits/folded-runs`'s `DefinitionProjectionMissingError` and
+ * `@corbits/workflows`'s `DefinitionProjectionMissingError` and
  * `MultiStepFoldUnsupportedError`. Its `message` is safe to return
  * verbatim; anything else is a platform fault whose internals must not
  * leak to a client.