[kitlangton]

With experimental.media_blob on, forking or importing a session now stores its images as blobs instead of copying the base64 inline.
Until now, a fork copied the parent's inline images into every child row, and an import wrote them inline too, so each copy carried the full image data. With the flag off, nothing changes, and existing rows are never rewritten.

❌ Before

1. A session's messages hold inline images (written before media blobs, or imported).
2. Fork it: each copied row ❌ carries the full base64 again.
3. Import a session: its images ❌ land inline in the new rows.

✅ After (flag on for that session's location)

1. Fork: the child's copied rows store each image once as a blob and point to it with a media: ref. The child records its own refs, so it still reads the same after the parent is deleted. The parent's rows stay as they were.
2. Import: the imported rows store their images as blobs, with refs, in the same transaction.
3. Flag off: forks and imports copy exactly as before.

  fork:   copy parent rows (parts merged inline) -> child
+         if media_blob: move inline media to blobs, store refs in the child row
  import: insert settled messages
+         if media_blob: move inline media to blobs, store refs, record blob refs

📊 Proof

Case	Before	After
Fork of an inline parent, flag on	❌ child rows inline	✅ child holds refs, parent untouched
Delete the parent, then read the fork	✅ inline copy	✅ child still reads the same from its own refs
Import with media, flag on	❌ inline	✅ refs plus blob refs, reads back unchanged
Flag off (fork and import)	✅ inline	✅ inline, no blobs written

The new fork and import tests fail without the change (checked by stashing projector.ts, the fork publish in session.ts, and transfer.ts). The server media test no longer has to convert imported rows to blobs itself.

👀 Read: Session.fork in session.ts and the fork projection. The parent's location decides whether the copy's media becomes blobs, and that one boolean reaches the projection through a context value. The caller's config, and its location, never decide it, and the published event stays as it was. A fork replayed without that value copies media inline, as before. Import reads the destination's config only when the import actually carries inline media. If a location can't load, the flag counts as off and media stays inline. The rest is routine.

🔁 Review: a cross-model review found four issues, all fixed with a test each:

1. A location that fails to boot raised a defect the old catch missed, crashing forks and imports even with the flag off. It now catches every failure except interrupts, logs a warning and keeps media inline.
2. The fork read the flag from the caller's config, and turning it on wrapped the publish in the parent's location, which changed where the event was routed. Now Session.fork resolves the flag once from the parent's location and passes only that boolean to the projection.
3. Import recorded refs for any media: link in the imported tool output, which could revive or expose a blob the import never held. It now records refs only for blobs taken out of its own inline data.
4. Import booted the destination's instance on every import. It now does that only when the import carries inline media; the server import test no longer needs model fetching turned off.

_{Part of t570 (media blobs plan, step 1) · no user reports · no data migration}

Checks run

• packages/core: bun test test/session-projector.test.ts test/session-prompt.test.ts test/session-create.test.ts test/bus-session-routing.test.ts: 135 pass, 0 fail
• packages/server: bun test test/media.test.ts test/session-import.test.ts: 7 pass, 0 fail
• bun typecheck in packages/core and packages/server: no errors
• Each review fix's test fails with that fix undone

---

Added by the opencode bot (AI-generated).

Related: #58436

[opencodejr]

Looks good. The change works as described: with experimental.media_blob on, forked and imported rows store their images as blobs and record the child's own refs. With the flag off, or when a location can't load, media stays inline. I put the base versions of projector.ts, transfer.ts and session.ts back and the new fork and import tests failed; with the change they pass. The related core and server suites pass and tsc is clean in core and server.

Public API: unchanged. The new SessionBlob exports are internal to core, the Forked event's payload is the same, and the experimental.media_blob key already existed.

No blocking concerns. One thing to keep in mind: a replayed fork copies media inline, as the description notes, so a replay can store the same rows differently, though reads come out the same.

Related: #58436