diff --git a/__tests__/lib/string.test.ts b/__tests__/lib/string.test.ts
index 8a56423796..da34bd0dc6 100644
--- a/__tests__/lib/string.test.ts
+++ b/__tests__/lib/string.test.ts
@@ -1,4 +1,4 @@
-import {RichText} from '@atproto/api'
+import {RichText} from '@bsky.app/sdk/richtext'
 import {i18n} from '@lingui/core'
 
 import {parseEmbedPlayerFromUrl} from '#/lib/strings/embed-player'
diff --git a/oxlint-suppressions.json b/oxlint-suppressions.json
index 6df56e5f78..6267907f16 100644
--- a/oxlint-suppressions.json
+++ b/oxlint-suppressions.json
@@ -1813,5 +1813,16 @@
     "typescript/no-unsafe-member-access": {
       "count": 3
     }
+  },
+  "webpack.config.js": {
+    "import/no-nodejs-modules": {
+      "count": 1
+    },
+    "typescript/no-unsafe-call": {
+      "count": 11
+    },
+    "typescript/no-unsafe-member-access": {
+      "count": 30
+    }
   }
 }
\ No newline at end of file
diff --git a/package.json b/package.json
index f58b0adcb7..281cb0d8b0 100644
--- a/package.json
+++ b/package.json
@@ -98,7 +98,6 @@
     "prettier": "prettier --check ."
   },
   "dependencies": {
-    "@atproto/api": "0.20.34",
     "@atproto/common-web": "0.5.6",
     "@atproto/lex": "0.2.2",
     "@atproto/lex-client": "0.2.4",
diff --git a/src/features/inviteFriends/urls.ts b/src/features/inviteFriends/urls.ts
index 7a22b8faf4..d75d5fa133 100644
--- a/src/features/inviteFriends/urls.ts
+++ b/src/features/inviteFriends/urls.ts
@@ -7,7 +7,7 @@
  * label simply drops the `https://` scheme for readability.
  *
  * Kept as a dependency-free leaf module (no #/lib/strings/url-helpers import)
- * so its unit tests stay fast and isolated from the heavy @atproto/api graph.
+ * so its unit tests stay fast and isolated from the heavy SDK graph.
  */
 
 function stripLeadingAt(handle: string): string {
diff --git a/src/lib/api/__tests__/computeCid.test.ts b/src/lib/api/__tests__/computeCid.test.ts
index 4082c6b553..a2cd6967f3 100644
--- a/src/lib/api/__tests__/computeCid.test.ts
+++ b/src/lib/api/__tests__/computeCid.test.ts
@@ -7,22 +7,20 @@
 jest.unmock('multiformats/cid')
 jest.unmock('multiformats/hashes/hasher')
 
-import {BlobRef} from '@atproto/api'
 import {CID} from 'multiformats/cid'
 
 import {computeCid} from '#/lib/api/computeCid'
 import {type app, type com} from '#/lexicons'
 
 /*
- * Golden-CID regression test for the composer post pipeline (design section F).
+ * Golden-CID regression test for the composer post pipeline.
  *
  * `computeCid` hashes a post record in the client so a thread's later posts can
  * reference earlier posts by CID before the server assigns them. The hash is
  * byte-sensitive: any drift in how records (especially blobs) are serialized to
  * DAG-CBOR silently produces the wrong CID and breaks reply chains with NO type
- * error. These golden values were captured from the PRE-migration `computeCid`
- * (the `instanceof BlobRef` path) and MUST remain byte-identical after the guard
- * is changed to the structural `isBlobRef` shape check.
+ * error. These golden values MUST remain byte-identical - they gate the
+ * structural `isBlobRef` shape check.
  *
  * The blob CID below is a fixed, deterministic CIDv1/raw/sha256 used purely as a
  * stable fixture - it is not derived from any real upload.
@@ -30,10 +28,7 @@ import {type app, type com} from '#/lexicons'
 const BLOB_CID = 'bafkreieq5jui4j25lacwomsqgjeswwl3y5zcdrresptwgmfylxo2depppq'
 
 /**
- * Build a post record with an image embed whose blob is the given value. Used to
- * prove that a `@atproto/api` `BlobRef` class instance (the shape the not-yet
- * -migrated video path still yields) and a plain-JSON lex blob (the shape lex
- * `uploadBlob` returns) hash to the SAME CID.
+ * Build a post record with an image embed whose blob is the given value.
  */
 function postWithImageBlob(blob: unknown): app.bsky.feed.post.Main {
   return {
@@ -65,22 +60,14 @@ describe('computeCid', () => {
     )
   })
 
-  it('case 2: record whose embed carries a BlobRef class instance', async () => {
-    const blob = new BlobRef(CID.parse(BLOB_CID), 'image/jpeg', 12345)
-    expect(await computeCid(postWithImageBlob(blob))).toBe(
-      'bafyreiem7g6vja66nebr7he4fshfnlyndyldbvle2n265oixscmepjcbii',
-    )
-  })
-
-  it('case 2b: a plain-JSON blob object hashes identically to the class instance', async () => {
+  it('case 2: record whose embed carries a plain-JSON lex blob', async () => {
     /*
-     * This is the post-migration shape: lex `uploadBlob` returns a plain object
-     * `{$type: 'blob', ref, mimeType, size}` (with `ref` a parsed CID), not a
-     * `BlobRef` class instance. The structural `isBlobRef` guard must treat it
-     * exactly like the class instance so the CID is unchanged. Under the
-     * pre-change `instanceof` code this case already matches because the plain
-     * object walks through `prepareForHashing` unchanged and DAG-CBOR encodes
-     * its CID `ref` the same way `.ipld()` does.
+     * The blob shape lex `uploadBlob` returns: a plain object
+     * `{$type: 'blob', ref, mimeType, size}` with `ref` a parsed CID. The
+     * structural `isBlobRef` guard passes it through `prepareForHashing`
+     * untouched and DAG-CBOR encodes its CID `ref` as a CID link. The golden
+     * CID below is the byte-identical value the pre-migration `BlobRef` class
+     * instance produced via `.ipld()`.
      */
     const blob = {
       $type: 'blob' as const,
diff --git a/src/lib/api/computeCid.ts b/src/lib/api/computeCid.ts
index 894f7a8c03..fdfb344e58 100644
--- a/src/lib/api/computeCid.ts
+++ b/src/lib/api/computeCid.ts
@@ -46,10 +46,9 @@ export async function computeCid(
 }
 
 /**
- * True for a plain-JSON lexicon blob, the shape lex `uploadBlob` now returns
- * (`{$type: 'blob', ref, mimeType, size}` with `ref` a parsed CID). Replaces
- * the old `instanceof BlobRef` check, since lex blobs are plain objects, not
- * class instances (design section F).
+ * True for a plain-JSON lexicon blob, the shape lex `uploadBlob` returns
+ * (`{$type: 'blob', ref, mimeType, size}` with `ref` a parsed CID). Lex blobs
+ * are plain objects, not class instances.
  */
 function isBlobRef(v: unknown): boolean {
   if (v == null || typeof v !== 'object') return false
@@ -57,24 +56,6 @@ function isBlobRef(v: unknown): boolean {
   return o.$type === 'blob' && 'ref' in o && 'mimeType' in o
 }
 
-/**
- * True for a legacy `@atproto/api` `BlobRef` class instance. During the
- * migration the video embed path still yields these (its blob comes from the
- * not-yet-migrated `app.bsky.video.getJobStatus` bridge call), so we must keep
- * handling them here even though the composer's own uploads are now plain lex
- * blobs. A class instance is duck-typed by its `ipld()` method plus the
- * `ref`/`mimeType` fields; it has NO `$type` and a non-plain prototype, so it
- * would otherwise slip past both `isBlobRef` and `isPlainObject` and be encoded
- * wrong - silently breaking video reply-chain CIDs.
- */
-function isBlobRefInstance(
-  v: unknown,
-): v is {ipld: () => unknown; ref: unknown; mimeType: unknown} {
-  if (v == null || typeof v !== 'object') return false
-  const o = v as Record<string, unknown>
-  return typeof o.ipld === 'function' && 'ref' in o && 'mimeType' in o
-}
-
 // Returns a transformed version of the object for use in DAG-CBOR.
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
 function prepareForHashing(v: any): any {
@@ -87,16 +68,6 @@ function prepareForHashing(v: any): any {
     return v
   }
 
-  /*
-   * A legacy `BlobRef` class instance must be converted via `ipld()` to the
-   * plain `{$type, ref, mimeType, size}` object; encoding the instance directly
-   * would emit its internal `original` field and omit `$type`, producing the
-   * wrong CID. `ipld()` returns exactly what `isBlobRef` accepts above.
-   */
-  if (isBlobRefInstance(v)) {
-    return v.ipld()
-  }
-
   // Walk through arrays
   if (Array.isArray(v)) {
     let pure = true
diff --git a/src/lib/api/index.ts b/src/lib/api/index.ts
index 1e50b1c635..305dbe1e5a 100644
--- a/src/lib/api/index.ts
+++ b/src/lib/api/index.ts
@@ -402,10 +402,9 @@ async function resolveMedia(
     return {
       $type: 'app.bsky.embed.video',
       /*
-       * The video blob is a legacy `@atproto/api` BlobRef from the not-yet
-       * -migrated video pipeline (getJobStatus, in composer state/video). Its
-       * structural shape matches the lexicon blob field; the CID hasher handles
-       * both class instances and plain lex blobs (see computeCid).
+       * The video blob is a plain lex blob from the video pipeline
+       * (getJobStatus, in composer state/video). Its structural shape matches
+       * the lexicon blob field and the CID hasher (see computeCid).
        */
       video: videoDraft.pendingPublish.blobRef,
       alt: videoDraft.altText || undefined,
diff --git a/src/lib/strings/__tests__/errors.test.ts b/src/lib/strings/__tests__/errors.test.ts
index aa9538032f..d8cff26985 100644
--- a/src/lib/strings/__tests__/errors.test.ts
+++ b/src/lib/strings/__tests__/errors.test.ts
@@ -1,4 +1,3 @@
-import {XRPCError} from '@atproto/api'
 import {
   type Procedure,
   type Query,
@@ -15,21 +14,9 @@ import {
 import {isErrorMaybeAppPasswordPermissions, isNetworkError} from '../errors'
 
 /**
- * Old-world fixture: `@atproto/api` XRPCError. `.headers` is a plain record,
- * `.error` is the lexicon code string, `.status` a numeric ResponseType enum.
- */
-function oldError(
-  status: number,
-  error: string,
-  headers?: Record<string, string>,
-) {
-  return new XRPCError(status, error, undefined, headers)
-}
-
-/**
- * New-world fixture: lex `XrpcResponseError`. Built from a WHATWG `Response`
- * (so `.headers` is a `Headers` object and `.status` is the numeric HTTP
- * status) plus a JSON error payload carrying `.error`.
+ * Lex `XrpcResponseError` fixture. Built from a WHATWG `Response` (so
+ * `.headers` is a `Headers` object and `.status` is the numeric HTTP status)
+ * plus a JSON error payload carrying `.error`.
  */
 function lexError(
   status: number,
@@ -45,8 +32,7 @@ function lexError(
 }
 
 describe('isXrpcError', () => {
-  it('matches both old XRPCError and lex XrpcResponseError', () => {
-    expect(isXrpcError(oldError(400, 'TokenInvalid'))).toBe(true)
+  it('matches a lex XrpcResponseError', () => {
     expect(isXrpcError(lexError(400, 'TokenInvalid'))).toBe(true)
   })
 
@@ -58,8 +44,7 @@ describe('isXrpcError', () => {
 })
 
 describe('getErrorStatus', () => {
-  it('reads status from both worlds', () => {
-    expect(typeof getErrorStatus(oldError(400, 'TokenInvalid'))).toBe('number')
+  it('reads status from a lex XrpcResponseError', () => {
     expect(getErrorStatus(lexError(429, 'RateLimitExceeded'))).toBe(429)
   })
 
@@ -69,8 +54,7 @@ describe('getErrorStatus', () => {
 })
 
 describe('getErrorName', () => {
-  it('reads the lexicon error code from both worlds', () => {
-    expect(getErrorName(oldError(400, 'TokenInvalid'))).toBe('TokenInvalid')
+  it('reads the lexicon error code', () => {
     expect(getErrorName(lexError(400, 'TokenInvalid'))).toBe('TokenInvalid')
   })
 
@@ -80,12 +64,6 @@ describe('getErrorName', () => {
 })
 
 describe('getErrorHeader', () => {
-  it('reads a header from the old record shape (case-insensitive)', () => {
-    const e = oldError(429, 'RateLimitExceeded', {'ratelimit-reset': '123'})
-    expect(getErrorHeader(e, 'ratelimit-reset')).toBe('123')
-    expect(getErrorHeader(e, 'RateLimit-Reset')).toBe('123')
-  })
-
   it('reads a header from the lex Headers object', () => {
     const e = lexError(429, 'RateLimitExceeded', {'ratelimit-reset': '123'})
     expect(getErrorHeader(e, 'ratelimit-reset')).toBe('123')
@@ -99,10 +77,7 @@ describe('getErrorHeader', () => {
 })
 
 describe('isErrorMaybeAppPasswordPermissions', () => {
-  it('matches a TokenInvalid error from both worlds', () => {
-    expect(
-      isErrorMaybeAppPasswordPermissions(oldError(400, 'TokenInvalid')),
-    ).toBe(true)
+  it('matches a TokenInvalid error', () => {
     expect(
       isErrorMaybeAppPasswordPermissions(lexError(400, 'TokenInvalid')),
     ).toBe(true)
@@ -118,9 +93,6 @@ describe('isErrorMaybeAppPasswordPermissions', () => {
   })
 
   it('does not match unrelated XRPC errors', () => {
-    expect(
-      isErrorMaybeAppPasswordPermissions(oldError(400, 'InvalidRequest')),
-    ).toBe(false)
     expect(
       isErrorMaybeAppPasswordPermissions(lexError(400, 'InvalidRequest')),
     ).toBe(false)
diff --git a/src/lib/strings/errors.ts b/src/lib/strings/errors.ts
index c737be01b3..7923fd7545 100644
--- a/src/lib/strings/errors.ts
+++ b/src/lib/strings/errors.ts
@@ -1,6 +1,6 @@
 import {t} from '@lingui/core/macro'
 
-import {getErrorName, getErrorStatus, isXrpcError} from '#/lib/xrpc-error'
+import {getErrorName, isXrpcError} from '#/lib/xrpc-error'
 
 export function cleanError(e: unknown): string {
   if (!e) {
@@ -87,13 +87,11 @@ export function isCancelledError(e: unknown) {
   return str.includes('cancel')
 }
 
-// TODO Replace this with error.shouldRetry() when available. -dsb
 const RETRYABLE_ERRORS = [408, 425, 429, 500, 502, 503, 504, 522, 524]
 export function isRetryableHttpStatus(status: number) {
   return RETRYABLE_ERRORS.includes(status)
 }
 
 export function shouldRetryError(e: unknown) {
-  const status = getErrorStatus(e)
-  return status !== undefined && isRetryableHttpStatus(status)
+  return isXrpcError(e) && e.shouldRetry()
 }
diff --git a/src/lib/xrpc-error.ts b/src/lib/xrpc-error.ts
index bc23fda70a..dd9e3ab9d3 100644
--- a/src/lib/xrpc-error.ts
+++ b/src/lib/xrpc-error.ts
@@ -1,44 +1,34 @@
-import {XRPCError} from '@atproto/api'
 import {XrpcError, XrpcResponseError} from '@atproto/lex-client'
 
 /**
- * True for an XRPC error from either the old bridge agent (`@atproto/api`
- * `XRPCError`) or a lex `Client` (`@atproto/lex-client` `XrpcError`, the
- * abstract base of `XrpcResponseError`/`XrpcInvalidResponseError`/
- * `XrpcInternalError`). During the migration both worlds can throw, so matchers
- * must accept both.
+ * True for an XRPC error from a lex `Client` (`@atproto/lex-client` `XrpcError`,
+ * the abstract base of `XrpcResponseError`/`XrpcInvalidResponseError`/
+ * `XrpcInternalError`).
  */
-export function isXrpcError(e: unknown): e is XRPCError | XrpcError {
-  return e instanceof XRPCError || e instanceof XrpcError
+export function isXrpcError(e: unknown): e is XrpcError {
+  return e instanceof XrpcError
 }
 
 /**
- * HTTP status, or undefined if not an XRPC error / no response. Only lex
+ * HTTP status, or undefined if not an XRPC error / no response. Only
  * `XrpcResponseError` (a server response) carries a status; the internal/fetch
  * lex errors do not.
  */
 export function getErrorStatus(e: unknown): number | undefined {
-  if (e instanceof XRPCError) return e.status
-  if (e instanceof XrpcResponseError) return e.status
-  return undefined
+  return e instanceof XrpcResponseError ? e.status : undefined
 }
 
-/** The lexicon error code (`err.error`), from either world. */
+/** The lexicon error code (`err.error`). */
 export function getErrorName(e: unknown): string | undefined {
-  if (isXrpcError(e)) return (e as {error?: string}).error
-  return undefined
+  return isXrpcError(e) ? (e as {error?: string}).error : undefined
 }
 
 /**
- * Read a response header off an XRPC error, normalizing the shape change:
- * old XRPCError.headers is a plain record; lex XrpcResponseError.headers is a
- * WHATWG Headers object.
+ * Read a response header off an XRPC error. Only `XrpcResponseError` carries a
+ * response; its `.headers` is a WHATWG `Headers` object.
  */
 export function getErrorHeader(e: unknown, name: string): string | undefined {
-  if (e instanceof XrpcResponseError) return e.headers.get(name) ?? undefined
-  if (e instanceof XRPCError) {
-    const h = (e as {headers?: Record<string, string>}).headers
-    return h?.[name.toLowerCase()]
-  }
-  return undefined
+  return e instanceof XrpcResponseError
+    ? (e.headers.get(name) ?? undefined)
+    : undefined
 }
diff --git a/src/screens/StarterPack/Wizard/StepProfiles.tsx b/src/screens/StarterPack/Wizard/StepProfiles.tsx
index 5d21328700..45e52fa14b 100644
--- a/src/screens/StarterPack/Wizard/StepProfiles.tsx
+++ b/src/screens/StarterPack/Wizard/StepProfiles.tsx
@@ -16,7 +16,6 @@ import {ScreenTransition} from '#/components/ScreenTransition'
 import {WizardProfileCard} from '#/components/StarterPack/Wizard/WizardListCard'
 import {Text} from '#/components/Typography'
 import {IS_NATIVE} from '#/env'
-import {type app} from '#/lexicons'
 import type * as bsky from '#/types/bsky'
 
 function keyExtractor(item: bsky.profile.AnyProfileView) {
diff --git a/src/state/messages/convo/agent.ts b/src/state/messages/convo/agent.ts
index 0c34f2c1d2..5427d72a3e 100644
--- a/src/state/messages/convo/agent.ts
+++ b/src/state/messages/convo/agent.ts
@@ -66,8 +66,8 @@ function toSystemMessageView(
    * The guard disjunction is kept inline in the `if` (rather than a separate
    * boolean) so TS narrows `ev` to the union of system-message log types, all
    * of which carry a `message: SystemMessageView`. The generated lexicon types
-   * are stricter than the old `@atproto/api` ones (no `[k: string]: unknown`
-   * index signature), so a separate boolean would not narrow the access.
+   * are strict (no `[k: string]: unknown` index signature), so a separate
+   * boolean would not narrow the access.
    */
   if (
     bsky.isType(chat.bsky.convo.defs.logAddMember, ev) ||
diff --git a/src/state/queries/explore-feed-previews.tsx b/src/state/queries/explore-feed-previews.tsx
index d1cc369473..4d1a36cfa3 100644
--- a/src/state/queries/explore-feed-previews.tsx
+++ b/src/state/queries/explore-feed-previews.tsx
@@ -205,7 +205,7 @@ export function useFeedPreviews(
               if (item.isFallbackMarker) continue
 
               const moderations = item.items.map(item =>
-                // @atproto/api and yields lex-typed slice items.
+                // the tuner yields lex-typed slice items.
                 moderatePost(item.post, moderationOpts!),
               )
 
@@ -232,7 +232,7 @@ export function useFeedPreviews(
                     const feedPostSliceItem: FeedPostSliceItem = {
                       _reactKey: `${item._reactKey}-${i}-${subItem.post.uri}`,
                       uri: subItem.post.uri,
-                      // off @atproto/api and yields lex-typed slice items.
+                      // the tuner yields lex-typed slice items.
                       post: subItem.post,
                       record: subItem.record,
                       moderation: moderations[i],
diff --git a/src/state/queries/post-feed.ts b/src/state/queries/post-feed.ts
index 0edb16faf2..d687ed7767 100644
--- a/src/state/queries/post-feed.ts
+++ b/src/state/queries/post-feed.ts
@@ -288,7 +288,7 @@ export function usePostFeedQuery(
                 .map(slice => {
                   const moderations = slice.items.map(item =>
                     moderatePost(
-                      // off @atproto/api and yields lex-typed slice items.
+                      // the tuner yields lex-typed slice items.
                       item.post,
                       moderationOpts!,
                     ),
@@ -341,7 +341,7 @@ export function usePostFeedQuery(
                       const feedPostSliceItem: FeedPostSliceItem = {
                         _reactKey: `${slice._reactKey}-${i}-${item.post.uri}`,
                         uri: item.post.uri,
-                        // off @atproto/api and yields lex-typed slice items.
+                        // the tuner yields lex-typed slice items.
                         post: item.post,
                         record: item.record,
                         moderation: moderations[i],
diff --git a/src/state/queries/search-posts-params.ts b/src/state/queries/search-posts-params.ts
index 33950edae1..78b549ce23 100644
--- a/src/state/queries/search-posts-params.ts
+++ b/src/state/queries/search-posts-params.ts
@@ -289,8 +289,8 @@ export function buildSearchPostsV2Filters(
 }
 
 /**
- * Consistent with timestamp parsing in @atproto/api. Only the date is used; the
- * time is appended here since the lexicon expects a datetime value.
+ * Consistent with atproto timestamp parsing. Only the date is used; the time
+ * is appended here since the lexicon expects a datetime value.
  */
 const parseTimestamp = (value: string | undefined): string | undefined => {
   if (!value) return undefined
diff --git a/src/types/bsky/__tests__/index.test.ts b/src/types/bsky/__tests__/index.test.ts
index f108ebc98a..44f0bd4aed 100644
--- a/src/types/bsky/__tests__/index.test.ts
+++ b/src/types/bsky/__tests__/index.test.ts
@@ -1,5 +1,3 @@
-import {AppBskyFeedPost} from '@atproto/api'
-
 import {app} from '#/lexicons'
 import * as bsky from '#/types/bsky'
 
@@ -44,8 +42,8 @@ describe('types/bsky new-world helpers (#/lexicons)', () => {
     })
 
     it('accepts an invalid body that has the right $type (dangerous semantics)', () => {
-      // Mirrors the old `dangerousIsType`: only the `$type` is checked, so a
-      // structurally invalid record still passes.
+      // Only the `$type` is checked, so a structurally invalid record still
+      // passes.
       expect(bsky.isType(app.bsky.feed.post, invalidPost)).toBe(true)
     })
 
@@ -55,8 +53,7 @@ describe('types/bsky new-world helpers (#/lexicons)', () => {
     })
 
     it('returns false (does not throw) for null and undefined', () => {
-      // Mirrors the old `dangerousIsType`/`is$typed` behavior - call sites
-      // pass e.g. `post.record` which may be undefined.
+      // Call sites pass e.g. `post.record` which may be undefined.
       expect(bsky.isType(app.bsky.feed.post, null)).toBe(false)
       expect(bsky.isType(app.bsky.feed.post, undefined)).toBe(false)
       expect(bsky.isType(app.bsky.feed.post, 'string')).toBe(false)
@@ -122,42 +119,3 @@ describe('types/bsky new-world helpers (#/lexicons)', () => {
     })
   })
 })
-
-describe('types/bsky legacy helpers (@atproto/api) still work', () => {
-  it('dangerousIsType accepts the right $type without validating the body', () => {
-    expect(
-      bsky.dangerousIsType<AppBskyFeedPost.Record>(
-        validPost,
-        AppBskyFeedPost.isRecord,
-      ),
-    ).toBe(true)
-    // Right $type, invalid body - still passes the fast guard.
-    expect(
-      bsky.dangerousIsType<AppBskyFeedPost.Record>(
-        invalidPost,
-        AppBskyFeedPost.isRecord,
-      ),
-    ).toBe(true)
-    expect(
-      bsky.dangerousIsType<AppBskyFeedPost.Record>(
-        wrongType,
-        AppBskyFeedPost.isRecord,
-      ),
-    ).toBe(false)
-  })
-
-  it('validate fully validates the body', () => {
-    expect(
-      bsky.validate<AppBskyFeedPost.Record>(
-        validPost,
-        AppBskyFeedPost.validateRecord,
-      ),
-    ).toBe(true)
-    expect(
-      bsky.validate<AppBskyFeedPost.Record>(
-        invalidPost,
-        AppBskyFeedPost.validateRecord,
-      ),
-    ).toBe(false)
-  })
-})
diff --git a/src/types/bsky/index.ts b/src/types/bsky/index.ts
index e5b292e3a2..278c1c1079 100644
--- a/src/types/bsky/index.ts
+++ b/src/types/bsky/index.ts
@@ -10,86 +10,21 @@ import {
   type ValidateOptions,
   type ValidationResult,
 } from '@atproto/lex'
-import {type ValidationResult as LegacyValidationResult} from '@atproto/lexicon'
 
 export * as post from '#/types/bsky/post'
 export * as profile from '#/types/bsky/profile'
 export * as starterPack from '#/types/bsky/starterPack'
 
-/**
- * Unsafe cast from an old `@atproto/api` view/record type to its `#/lexicons`
- * equivalent. Only for mixed-world boundaries during the migration where a
- * producer has not yet flipped. Structurally the shapes are identical modulo
- * branded string types; this asserts the brand the compiler cannot prove.
- * Every use is a migration debt marker - grep `toLex` at Phase 4 cleanup.
- */
-export function toLex<T>(v: unknown): T {
-  return v as T
-}
-
-/**
- * Fast type checking without full schema validation, for use with data we
- * trust, or for non-critical path use cases. Why? Our SDK's `is*` identity
- * utils do not assert the type of the entire object, only the `$type` string.
- *
- * For full validation of the object schema, use the `validate` export from
- * this file.
- *
- * Usage:
- * ```ts
- * import * as bsky from '#/types/bsky'
- *
- * if (bsky.dangerousIsType<AppBskyFeedPost.Record>(item, AppBskyFeedPost.isRecord)) {
- *   // `item` has type `$Typed<AppBskyFeedPost.Record>` here
- * }
- * ```
- */
-export function dangerousIsType<R extends {$type?: string}>(
-  record: unknown,
-  identity: <V>(v: V) => v is V & {$type: NonNullable<R['$type']>},
-): record is R {
-  return identity(record)
-}
-
-/**
- * Fully validates the object schema, which has a performance cost.
- *
- * For faster checks with data we trust, like that from our app view, use the
- * `dangerousIsType` export from this same file.
- *
- * Usage:
- * ```ts
- * import * as bsky from '#/types/bsky'
- *
- * if (bsky.validate(item, AppBskyFeedPost.validateRecord)) {
- *   // `item` has type `$Typed<AppBskyFeedPost.Record>` here
- * }
- * ```
- */
-export function validate<R extends {$type?: string}>(
-  record: unknown,
-  validator: (v: unknown) => LegacyValidationResult<R>,
-): record is R {
-  return validator(record).success
-}
-
 /*
- * New-world helpers (below) operate on the generated lexicon schema objects
- * from '#/lexicons', replacing the `@atproto/api`-based helpers above.
- *
- * The old helpers take a record plus a standalone `is*`/`validate*` function
- * (e.g. `AppBskyFeedPost.isRecord`). The new codegen instead attaches the
- * validation surface directly to each schema, so these helpers take the schema
- * object itself:
+ * These helpers operate on the generated lexicon schema objects from
+ * '#/lexicons'. The codegen attaches the validation surface directly to each
+ * schema, so these helpers take the schema object itself:
  *
  * ```ts
  * import {app} from '#/lexicons'
  * import * as bsky from '#/types/bsky'
  *
- * // old: bsky.dangerousIsType(v, AppBskyFeedPost.isRecord)
  * bsky.isType(app.bsky.feed.post, v)
- *
- * // old: bsky.validate(v, AppBskyFeedPost.validateRecord)
  * bsky.matches(app.bsky.feed.post, v)
  * ```
  *
@@ -113,9 +48,8 @@ type TypedSchema = RecordSchema | TypedObjectSchema
  * string; it does NOT assert the rest of the object matches the schema. An
  * invalid record with the right `$type` will pass.
  *
- * This is the '#/lexicons' equivalent of {@link dangerousIsType}. For full
- * validation of the object schema, use {@link matches}, {@link parse}, or
- * {@link safeParse} from this same file.
+ * For full validation of the object schema, use {@link matches}, {@link parse},
+ * or {@link safeParse} from this same file.
  *
  * Usage:
  * ```ts
@@ -135,10 +69,9 @@ export function isType<S extends TypedSchema>(
    * Deliberately NOT delegating to the schema's `isTypeOf`: the generated
    * `TypedObjectSchema.isTypeOf` treats a MISSING `$type` as a match
    * (maybe-typed semantics), which would let any plain object satisfy any
-   * def-schema check and break `$type`-discriminated unions. The old
-   * `dangerousIsType` required a present, matching `$type`, and so do we.
-   * The nullish/object guard also mirrors the old `is$typed` behavior of
-   * returning false (not throwing) for null/undefined input.
+   * def-schema check and break `$type`-discriminated unions. We require a
+   * present, matching `$type`. The nullish/object guard also returns false
+   * (not throws) for null/undefined input.
    */
   return (
     value != null &&
@@ -151,9 +84,8 @@ export function isType<S extends TypedSchema>(
  * Fully validates the object against the schema (strict, no coercion), which
  * has a performance cost, and narrows the value on success.
  *
- * This is the '#/lexicons' equivalent of {@link validate}. For faster checks
- * with data we trust, like that from our app view, use {@link isType} from this
- * same file.
+ * For faster checks with data we trust, like that from our app view, use
+ * {@link isType} from this same file.
  *
  * Usage:
  * ```ts
diff --git a/src/types/bsky/starterPack.ts b/src/types/bsky/starterPack.ts
index ed080864ab..21cb4634df 100644
--- a/src/types/bsky/starterPack.ts
+++ b/src/types/bsky/starterPack.ts
@@ -1,11 +1,10 @@
 import {app} from '#/lexicons'
 
 /*
- * The generated `$type`-only guards. The old `@atproto/api`
- * `AppBskyGraphDefs.isStarterPackView*` helpers matched on a present,
- * matching `$type`; we reproduce that here against the `#/lexicons` schema's
- * `$type` string rather than delegating to the schema's `isTypeOf` (which
- * treats a missing `$type` as a match).
+ * The generated `$type`-only guards. These match on a present, matching
+ * `$type` against the `#/lexicons` schema's `$type` string rather than
+ * delegating to the schema's `isTypeOf` (which treats a missing `$type` as a
+ * match).
  */
 export function isBasicView(
   v: unknown,
diff --git a/src/view/com/composer/drafts/state/schema.ts b/src/view/com/composer/drafts/state/schema.ts
index 123fc1d715..7a4f4c44f7 100644
--- a/src/view/com/composer/drafts/state/schema.ts
+++ b/src/view/com/composer/drafts/state/schema.ts
@@ -1,7 +1,7 @@
 import {type app} from '#/lexicons'
 /**
  * Types for draft display and local media tracking.
- * Server draft types come from @atproto/api.
+ * Server draft types come from the atproto lexicons.
  */
 
 /**
diff --git a/webpack.config.js b/webpack.config.js
index 889a2cb83f..dace4e8fe5 100644
--- a/webpack.config.js
+++ b/webpack.config.js
@@ -45,7 +45,7 @@ function patchSourceMapFilter(rules, pathPattern) {
 
 module.exports = async function (env, argv) {
   env.babel = {
-    dangerouslyAddModulePathsToTranspile: ['@bsky.app/expo', '@atproto/api'],
+    dangerouslyAddModulePathsToTranspile: ['@bsky.app/expo'],
   }
   let config = await createExpoWebpackConfigAsync(env, argv)
   /*