fix(core,sdk): stop chat.agent losing messages during recovery - #4907

…ptStorage seam

Introduces the TranscriptStorage interface (load/save over id-addressed changes: put, remove, truncateAfter, state) and makes the built-in snapshot writer its default implementation. The runtime keeps a shadow of the transcript it last saved and hands the storage the diff after every turn, failed turn and history-changing action, together with the stream cursors the next boot resumes from. The default storage reduces each changeset onto an in-memory copy and rewrites the blob as version 2, so a turn still costs one PUT and no GET, and the boot read goes through load.

The snapshot read/write helpers move to their own module so the storage can import them without a cycle; the test seams keep their import path. The mock chat agent harness now reports version 2 snapshots and accepts either version as a seed.

…ript storage state

The model lane after a compaction cannot be rebuilt from the transcript, so every continuation used to re-read the whole conversation and summarise it again. The runtime now records the compacted lane in the storage's state slot, with the transcript id it covers and a fingerprint of that prefix, and rebuilds from it at boot when the prefix is unchanged. A rollback or edit that reconverts the lane clears the state in the same changeset as the truncate.

Conversational messages added with chat.inject are recorded the same way, anchored to the transcript message they followed, so they survive a continuation instead of living only in the worker that received them.

Adds an in-memory storage that logs the changesets it receives, and a test-only override for the storage the runtime persists through, so the exact changesets for a turn, a mid-turn steer, a compaction, a rollback and an injection are asserted.

…t storage own the model's context

One condition used to decide three things at boot: whether to read the persisted transcript, whether to replay the session's output tail, and whether to replay unacknowledged input. Registering hydrateMessages switched all three off, so an app that owned its own context also lost crash recovery, and no application can rebuild the tail its dead run had already emitted. The replays and onRecoveryBoot now run for every agent; only the transcript read is skipped for hydrateMessages.

The storage can now declare loadContext, which the runtime calls on every turn and action in place of the accumulated transcript, the role hydrateMessages played, while save keeps receiving every change. hydrateMessages is deprecated with a one-time warning, and configuring it together with a storage that has loadContext is an error.

…nformance suite

chat.agent takes a storage option, with the platform snapshot as the default and the TranscriptStorage types, defaultStorage, snapshotTranscriptStorage, memoryTranscriptStorage and reduceTranscriptChanges exported from @trigger.dev/sdk/ai. chat.createLoadTranscriptAction(storage) reads a conversation on the server for any storage, and useLoadTranscript renders it in the browser and seeds the transport's resume cursor. runTranscriptStorageTests from @trigger.dev/sdk/ai/test is the contract a storage implementation has to meet.

A new secret-key endpoint, GET /api/v1/sessions/:id/transcript, pages the platform snapshot server-side so rendering the last page of a long conversation does not download all of it; the default storage uses it for paged reads and falls back to the whole blob.

… message rows

The real-schema conformance target for the storage contract: one row per message in chat_messages, with the runtime's state and cursors in two new nullable columns on chats. Paging is by position in SQL. The agent itself still persists through its hooks.

…, and the changeset

Treat a non-positive page limit as no limit instead of an empty page a
caller cannot tell apart from the end of the transcript. Hold a seeded
resume cursor until the session exists, so a transcript load that resolves
before the session is created still opens the live stream past the
persisted history instead of replaying it.

Consume a pending resume cursor in setSession() and stop seedResumeCursor
from moving an existing session's cursor backward, so a transcript load that
resolves around session creation, or after the live stream already advanced,
does not replay persisted events. Add conformance coverage for a partial
answer's non-final status, and drive the resume-cursor unit tests through a
real transport. Docs: loadContext also fires on actions, a custom storage is
not limited to row-per-message backends, and the bring-your-own example
imports anthropic.