## Goal Serve Duo Workflow checkpoint `channel_values` by reconstructing them from **incremental blobs** instead of loading the full checkpoint header JSONB. This is the **read** half of the incremental-checkpoint work (parent epic [&22628](https://gitlab.com/groups/gitlab-org/-/work_items/22628)); the write path (phase 1) is tracked in [604684](https://gitlab.com/gitlab-org/gitlab/-/work_items/604684). The read machinery (`Gitlab::DuoWorkflow::ChannelValuesReconstructor` + `Ai::DuoWorkflows::Workflow` read methods) is shared; each **consumer** switches from the header read to blob reconstruction independently. ## Feature-flag model - **Master kill switch**: `duo_workflow_read_incremental_checkpoints` — disables reconstruction for every consumer at once. Rollout issue [604687](https://gitlab.com/gitlab-org/gitlab/-/work_items/604687). - **Per-consumer flags**: each consumer is additionally gated by its own flag so it can be rolled out and rolled back independently. A consumer reconstructs from blobs only when **both** the kill switch **and** its own flag are on. ### Gateway: two read entry points The gateway loads checkpoints through **two** paths, gated by **different** per-consumer flags: - **Normal resume / restart-continue** (START / RESUME / RETRY — the dominant path): LangGraph calls `aget_tuple` with no `checkpoint_id` and loads the latest checkpoint via the GraphQL `latestCheckpoint { compressedCheckpoint }` field. That field reconstructs from blobs through `WorkflowCheckpointEventPresenter#checkpoint` — the raw-checkpoint GraphQL consumer, gated by `dw_read_blobs_graphql` (!247391). - **Stop / interruption recovery** (`STOP_RECOVERY`): the flow walks the checkpoint chain to a boundary (newest `INPUT_REQUIRED` pause, or session start) and pins that `checkpoint_id`; LangGraph then loads that specific checkpoint via the internal `by_thread_ts` endpoint, gated by `dw_read_blobs_api` (!247134 + gateway ai-assist!6363). For the `by_thread_ts` path the read decision lives in the gateway (AIGW), which gates on the kill switch **and** `dw_read_blobs_api` **and** the workflow's `incremental_checkpoints_enabled` column, falling back to the full-checkpoint read when either flag is off (no state loss). The Rails `by_thread_ts` endpoint serves on header existence. Note the `latestCheckpoint` path still _sources_ its header record from `checkpoints.latest` (full table) and overlays reconstructed channels onto it — see #605653 for retiring that table. ## Deliverables Per-flag MR, rollout issue, and current state are tracked in the rollout table of the parent epic [&22628](https://gitlab.com/groups/gitlab-org/-/work_items/22628) (no separate list here); per-consumer detail lives in this epic's child issues. The GraphQL fields collapse to a single `dw_read_blobs_graphql` flag because they share the same presenter/query surface and cannot be rolled out independently of each other. ## Consumer inventory (source of truth) | Consumer | Reads | Where | Own flag | |---|---|---|---| | Gateway — normal resume (`latestCheckpoint`) | all channels | `latestCheckpoint`/`firstCheckpoint` → `WorkflowCheckpointEventPresenter#checkpoint` / `compressed_checkpoint` (header from `checkpoints.latest`, channels overlaid from blobs) | `dw_read_blobs_graphql` (shared) | | Gateway — stop-recovery (`by_thread_ts`) | all channels | `workflows_internal.rb` `GET by_thread_ts` (reconstructs); legacy `GET checkpoints` list/`278964` still read full table | `dw_read_blobs_api` | | Gateway — internal checkpoint list (`checkpoints_reversed`, pre-18.8 fallback) | all channels | `workflows_internal.rb` `GET checkpoints` | `dw_read_blobs_list` | | Trace download | ui_chat_log/plan/status | `workflows.rb` `trace_channel_values` | `dw_read_blobs_trace` | | Session messages GraphQL | ui_chat_log | `WorkflowCheckpointEventPresenter#duo_messages`/`#last_duo_message` | `dw_read_blobs_graphql` | | `executionStatus` GraphQL | status | `workflow_event_type.rb` → presenter `execution_status` | `dw_read_blobs_graphql` (shared) | | Messaging / notifications | ui_chat_log | `checkpoint.rb#ui_chat_log` → callback worker, progress reader, result email | `dw_read_blobs_notifications` | | Goal backfill | `__start__.goal` | `create_checkpoint_service` — reads the checkpoint being written | — |