agents: replay event log from Run.conversation() on worker crash (cursor only) #956

New issue

Closed

opened 2026-05-08 12:06:07 +00:00 by claude-desktop · 3 comments

claude-desktop commented

2026-05-08 12:06:07 +00:00

Collaborator

User story

As an operator whose service occasionally hard-restarts mid-task, I want cursor-backed sessions to recover their event timeline from Run.conversation() rather than being a black hole between restart and the next operator action, so I don't lose visibility into what the agent already did.

Context

@cursor/sdk exposes Run.conversation(): Promise<ConversationTurn[]> — the structured per-turn history of a run. Today after a crash + restart, redispatch-interrupted re-queues the task, but the old event log is gone (in-memory ring buffer + nothing rebuilt). For claude-code there's no equivalent fallback (the SDK doesn't expose a transcript replay endpoint), so this issue scopes to cursor only.

Acceptance criteria

Cursor adapter

On Agent.resume(...) — when there is a stored cursor session — call Run.conversation() against the most recent run for that agent and translate each ConversationTurn into the existing TaskEvent types: assistant, user, tool_progress, tool_summary. Yield them at the head of the run's event stream so the operator immediately sees the prior context.
Mark the replayed events with a flag (replayed: true) so the event-log + UI can render them differently (faded / dimmed / "from previous run" divider).
Never re-execute side effects — replay is read-only; the runner skips PR-URL extraction (already extracted last time) and skips session-id capture for replayed events.

Event log

New TaskRecord.events[i].replayed: boolean field. Default false.
SSE envelope carries the flag.

Frontend

Replayed events render with reduced opacity + a single "Replayed from session X" divider above the run.

Tests

Unit test: feed a fake Run.conversation() result; assert correct event-stream prefix.
Integration test against a real cursor SDK using a recorded fixture run (no live API calls in CI).
Idempotence test: replay called twice does not duplicate events in the record.

Out of scope

Claude Code replay — no SDK support.
Replay for completed (non-resumed) runs — only fires when an agent is resumed mid-flight.

References

Parent: #950
Cursor SDK: Run.conversation() in node_modules/.bun/@cursor+sdk@1.0.12/node_modules/@cursor/sdk/dist/cjs/run.d.ts
Existing redispatch path: just redispatch-interrupted

## User story As an operator whose service occasionally hard-restarts mid-task, I want cursor-backed sessions to recover their event timeline from `Run.conversation()` rather than being a black hole between restart and the next operator action, so I don't lose visibility into what the agent already did. ## Context `@cursor/sdk` exposes `Run.conversation(): Promise<ConversationTurn[]>` — the structured per-turn history of a run. Today after a crash + restart, `redispatch-interrupted` re-queues the task, but the old event log is gone (in-memory ring buffer + nothing rebuilt). For claude-code there's no equivalent fallback (the SDK doesn't expose a transcript replay endpoint), so this issue scopes to cursor only. ## Acceptance criteria ### Cursor adapter - [ ] On `Agent.resume(...)` — when there is a stored cursor session — call `Run.conversation()` against the most recent run for that agent and translate each `ConversationTurn` into the existing `TaskEvent` types: `assistant`, `user`, `tool_progress`, `tool_summary`. Yield them at the head of the run's event stream so the operator immediately sees the prior context. - [ ] Mark the replayed events with a flag (`replayed: true`) so the event-log + UI can render them differently (faded / dimmed / "from previous run" divider). - [ ] Never re-execute side effects — replay is read-only; the runner skips PR-URL extraction (already extracted last time) and skips session-id capture for replayed events. ### Event log - [ ] New `TaskRecord.events[i].replayed: boolean` field. Default `false`. - [ ] SSE envelope carries the flag. ### Frontend - [ ] Replayed events render with reduced opacity + a single "Replayed from session X" divider above the run. ### Tests - [ ] Unit test: feed a fake `Run.conversation()` result; assert correct event-stream prefix. - [ ] Integration test against a real cursor SDK using a recorded fixture run (no live API calls in CI). - [ ] Idempotence test: replay called twice does not duplicate events in the record. ## Out of scope - Claude Code replay — no SDK support. - Replay for completed (non-resumed) runs — only fires when an agent is resumed mid-flight. ## References - Parent: #950 - Cursor SDK: `Run.conversation()` in `node_modules/.bun/@cursor+sdk@1.0.12/node_modules/@cursor/sdk/dist/cjs/run.d.ts` - Existing redispatch path: `just redispatch-interrupted`

claude-desktop added the

area:agents

type:user-story

labels

2026-05-08 12:07:01 +00:00

claude-desktop referenced this issue

2026-05-08 12:15:22 +00:00

dashboard: subagent tree (rail) + swimlane (timeline) for parallel agent runs #965

claude-desktop added a new dependency

2026-05-08 12:19:19 +00:00

#951 agents: stream cursor InteractionUpdate deltas (text, thinking, tool, shell, token, summary)

claude-desktop referenced this issue