[Bug]: Cursor shell commands never stream live output

Before submitting

I searched existing issues and did not find a duplicate.
I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

Use the Cursor provider (@cursor/sdk 1.2.0) and start a turn.
Ask the agent to run a command that prints over several seconds, e.g. for i in 1 2 3 4 5; do echo "tick $i"; sleep 2; done.
Expand the command row while the command is still running.

Expected behavior

Output appears incrementally while the command runs, the same way it does for Codex and Claude Agent. The expanded row should show tick 1, tick 2, … as they are printed.

Actual behavior

The expanded command row stays empty for the whole run. The complete output appears at once when the tool call finishes, so a long command looks hung.

This affects every Cursor shell command, including a real ~3 minute remote deploy where the whole log landed only at completion.

Impact

Major degradation or frequent failure

Version or commit

main @ cefb544 (also reproduced on 0.1.7-nightly.20261103.4119)

Environment

Linux, web client, Cursor provider via @cursor/sdk 1.2.0, default Cursor model, local mode.

Logs or stack traces

Redacted reproduction log. The command emits 5 lines over ~10s; T3 shows nothing until the end.

[12:00:00] INFO  running command
[12:00:10] INFO  command finished (5 lines of stdout)

# T3 UI while running: command row expanded, no output
# T3 UI after completion: all 5 lines appear at once

Root cause is in the adapter, not the transport. Cursor sends the delta as a protobuf oneof:

{ "case": "stdout", "value": { "data": "tick 1\n" } }
{ "case": "stderr", "value": { "data": "" } }
{ "case": "exit",   "value": { "code": 0, "aborted": false } }

packages/provider-cursor/src/server/adapter.ts (shellOutputText, ~L1975) only looks at
event.text, event.output, event.data, event.stdout, event.stderr, event.chunk, so it
returns "". shell-output-delta (~L2048) then returns early at:

const chunk = shellOutputText(update.event);
if (chunk.length === 0) return;

So command_output.delta is never emitted for Cursor, while the final output still arrives through
the completion fallback in emitToolArtifacts (cursorToolOutputText(toolCall) || streamedOutput).

The SDK does stream live: @cursor/sdk dist/esm/336.js runs handleShellStream concurrently with
the tool call and sends one shellOutputDelta update per stdout/stderr chunk.

Workaround

Run the command outside T3 (terminal/ssh) and paste the result, or keep the row collapsed and wait
for the completion output.

Proposed implementation plan

Posted as the working plan for this issue. No code is implemented yet.

Cursor live shell output — Implementation Plan

Execute task by task. Each task is one TDD cycle and one commit.

Goal: Cursor shell tool calls emit command_output.delta while the command runs, so an
expanded command row streams output live instead of dumping everything on completion.

Architecture: @cursor/sdk already emits one shell-output-delta interaction update per
stdout/stderr chunk, shaped { case: "stdout" | "stderr", value: { data } }. T3 receives every
update (adapter.ts → onDelta → handleInteractionUpdate) but shellOutputText reads flat keys,
returns "", and adapter.ts:2055 drops the chunk. Fix the extractor; the hub, RPC, capability,
and UI already work.

Tech Stack: TypeScript, Effect, @effect/vitest, Cursor replay harness, vp.

Constraints and non-goals

No changes to @cursor/sdk or node_modules. It is closed source and already correct.
No packages/contracts, apps/server, or apps/web changes. CommandOutputHub,
subscribeCommandOutput, and CommandOutputPanel already stream for Codex/Claude.
Nested subagent shells stay unsupported: the SDK typings say shell deltas are not exposed for
nested task updates (NestedTaskUpdateSchema comment).
Collapsed rows stay silent; that subscription behavior is unchanged.
Live output must not be persisted into the event store. The fixture asserts it.
exit payloads ({ code, aborted }) carry no text and must be ignored.

Invariants the fix must preserve

chunk is still appended to shell.streamedOutput (bounded tail, used as completion fallback).
One command_output.delta event per non-empty chunk, with the same itemId derivation.
Final persisted output still comes from cursorToolOutputText(toolCall) || streamedOutput.
No behavior change for non-shell tools or for Codex/Claude adapters.

Baseline (must fail before the fix)

vp test run apps/server/src/orchestration-v2/testkit/CursorReplayFixtures.integration.test.ts \
  -t "command_output_streaming/cursor"

Expected on main: FAIL — the running command must stream its output.

Task 1 — Add the failing Cursor replay fixture

Files:

Create: apps/server/src/orchestration-v2/testkit/fixtures/command_output_streaming/cursor_transcript.ndjson
Modify: apps/server/src/orchestration-v2/testkit/fixtures/command_output_streaming/output.ts:17-64
Modify: apps/server/src/orchestration-v2/testkit/fixtures/index.ts:1883-1905

Step 1: Author the transcript

The replay harness accepts hand-authored frames; model it on
fixtures/turn_interrupt_mid_tool/cursor_transcript.ndjson. Required order:

{"type":"transcript_start","provider":"cursor","protocol":"cursor-agent-sdk.local","version":"1","scenario":"command_output_streaming","metadata":{"generatedBy":"hand-authored","nativeAgentId":"agent-command-output-streaming"}}
{"type":"expect_outbound","label":"agent.open","frame":{"type":"agent.open","operation":"create","options":{"model":{"id":"composer-2.5"},"hasName":true,"mode":"agent","local":{"hasCwd":true,"autoReview":false,"settingSources":["project","user","team","mdm","plugins"],"sandboxEnabled":true,"enableAgentRetries":true}}}}
{"type":"emit_inbound","label":"agent.opened","frame":{"type":"agent.opened","agentId":"agent-command-output-streaming"}}
{"type":"expect_outbound","label":"run.start:1","frame":{"type":"run.start","message":"<the COMMAND_OUTPUT_STREAMING_PROMPT text>","options":{"model":{"id":"composer-2.5"},"mode":"agent"}}}
{"type":"emit_inbound","label":"run.started:1","frame":{"type":"run.started","runId":"run-command-output-streaming","agentId":"agent-command-output-streaming"}}
{"type":"emit_inbound","label":"tool-call-started","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"tool-call-started","callId":"tool_command_output_streaming","toolCall":{"type":"shell","args":{"command":"printf 'cursor line 1\\n'; sleep 1; printf 'cursor line 2\\n'; sleep 1; printf '\\033[0;32mcursor line 3\\033[0m\\n'; sleep 1; printf 'cursor line 4\\nprogress 100%\\n'","timeout":30000}},"modelCallId":"model-call-command-output-streaming"}}}
{"type":"emit_inbound","label":"shell-stdout-1","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"shell-output-delta","event":{"case":"stdout","value":{"data":"cursor line 1\n"}}}}}
{"type":"emit_inbound","label":"shell-stdout-2","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"shell-output-delta","event":{"case":"stdout","value":{"data":"cursor line 2\n"}}}}}
{"type":"emit_inbound","label":"shell-stdout-3","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"shell-output-delta","event":{"case":"stdout","value":{"data":"\u001b[0;32mcursor line 3\u001b[0m\n"}}}}}
{"type":"emit_inbound","label":"shell-stdout-4","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"shell-output-delta","event":{"case":"stdout","value":{"data":"cursor line 4\nprogress 100%\n"}}}}}
{"type":"emit_inbound","label":"shell-exit","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"shell-output-delta","event":{"case":"exit","value":{"code":0,"aborted":false}}}}}
{"type":"emit_inbound","label":"tool-call-completed","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"tool-call-completed","callId":"tool_command_output_streaming","toolCall":{"type":"shell","args":{"command":"..."},"result":{"status":"success","value":{"exitCode":0,"signal":"","stdout":"cursor line 1\ncursor line 2\ncursor line 3\ncursor line 4\nprogress 100%\n","stderr":"","executionTime":4000}}},"modelCallId":"model-call-command-output-streaming"}}}
{"type":"emit_inbound","label":"turn-ended","frame":{"type":"interaction.update","runId":"run-command-output-streaming","update":{"type":"turn-ended"}}}
{"type":"emit_inbound","label":"run.completed:1","frame":{"type":"run.completed","result":{"id":"run-command-output-streaming","requestId":"model-call-command-output-streaming","status":"finished","model":{"id":"composer-2.5"},"durationMs":4500}}}
{"type":"expect_outbound","label":"agent.close","frame":{"type":"agent.close","agentId":"agent-command-output-streaming"}}

Use the exact text of COMMAND_OUTPUT_STREAMING_PROMPT from
apps/server/src/orchestration-v2/testkit/fixtures/shared.ts in run.start.message, otherwise the
user-message assertion fails.

Step 2: Parameterize the assertion

In output.ts, extract the provider-neutral body into assertCommandOutputStreamingLines(input)
with fields { result, transcript, liveLines, liveAnsi?, excludedLiveLines? }. Keep two exports:

export function assertCommandOutputStreamingOutput(result, transcript) {
  assertCommandOutputStreamingLines({
    result,
    transcript,
    liveLines: ["stream line 2", "stream line 3", "stream line 4", "progress 100%"],
    liveAnsi: "\u001b[0;32mstream line 2\u001b[0m",
    excludedLiveLines: ["progress 50%"],
  });
}

export function assertCursorCommandOutputStreamingOutput(result, transcript) {
  assertCommandOutputStreamingLines({
    result,
    transcript,
    liveLines: ["cursor line 1", "cursor line 2", "cursor line 3", "cursor line 4", "progress 100%"],
    liveAnsi: "\u001b[0;32mcursor line 3\u001b[0m",
  });
}

Move the "Codex omits pre-yield command output" comment into the Codex wrapper.

Step 3: Register the provider

In fixtures/index.ts, add to command_output_streaming.providers:

{
  driver: ProviderDriverKind.make("cursor"),
  transcriptFile: new URL(
    "./command_output_streaming/cursor_transcript.ndjson",
    import.meta.url,
  ),
  modelSelection: CURSOR_MODEL_SELECTION,
  assertOutput: assertCursorCommandOutputStreamingOutput,
},

Also import assertCursorCommandOutputStreamingOutput from ./command_output_streaming/output.ts.

Step 4: Verify it fails for the right reason

vp test run apps/server/src/orchestration-v2/testkit/CursorReplayFixtures.integration.test.ts \
  -t "command_output_streaming/cursor"

Expected: FAIL at the running command must stream its output (chunks === 0).

Step 5: Commit

git add apps/server/src/orchestration-v2/testkit/fixtures/command_output_streaming/ \
        apps/server/src/orchestration-v2/testkit/fixtures/index.ts
git commit -m "test(server): cover Cursor live shell output streaming"

Task 2 — Read the Cursor `{ case, value }` payload

Files:

Modify: packages/provider-cursor/src/server/adapter.ts:1975-1984

Step 1: Replace the extractor

const shellOutputText = (event: Record<string, unknown>): string => {
  // The SDK flattens a protobuf oneof to { case: "stdout" | "stderr" | "exit", value }.
  const kind = event.case;
  if (kind === "stdout" || kind === "stderr") {
    const value = event.value;
    const data =
      typeof value === "object" && value !== null && "data" in value
        ? (value as { data?: unknown }).data
        : undefined;
    return typeof data === "string" ? data : "";
  }
  if (typeof kind === "string") {
    return "";
  }
  const candidates = [event.text, event.output, event.data, event.stdout, event.stderr, event.chunk];
  return candidates.filter((value): value is string => typeof value === "string").join("");
};

The trailing flat-key path stays as a fallback for SDK builds that emit unwrapped events.

Step 2: Run the fixture

vp test run apps/server/src/orchestration-v2/testkit/CursorReplayFixtures.integration.test.ts \
  -t "command_output_streaming"

Expected: PASS for command_output_streaming/cursor, and the existing
command_output_streaming/codex + command_output_streaming/claudeAgent still pass.

Step 3: Commit

git add packages/provider-cursor/src/server/adapter.ts
git commit -m "fix(provider-cursor): stream live shell output"

Task 3 — Verification

Run these, in order:

vp test run apps/server/src/orchestration-v2/testkit/CursorReplayFixtures.integration.test.ts
vp test run packages/provider-cursor/src/server/adapter.test.ts
vp run check:changed

Expected: all pass; check:changed reports no format, lint, knip, or typecheck errors.

Manual check (one real Cursor turn, cheapest model):

Start a dev server and open a Cursor thread.
Ask it to run: for i in 1 2 3 4 5; do echo "tick $i"; sleep 2; done
Expand the command row while it runs.
Expect tick 1…tick 5 to appear incrementally, then the completed output to remain.

If only the final output appears, the transcript test passed for the wrong reason: confirm the
event.case/event.value.data shape in the provider event log before changing anything else.

Risks

Risk	Mitigation
SDK version drift changes the oneof shape	Flat-key fallback retained; fixture pins the current shape
`exit` chunk appended as text	Explicit `kind === "string"` early return; fixture includes an `exit` frame
Duplicate output (live tail + final artifact)	Existing hub/final split is unchanged; fixture asserts running items carry no `output`
Subagent shells appear broken after the fix	Out of scope by SDK design; document in the PR as not covered

Rollback

Revert the two commits. No migrations, contracts, or persisted state change.

Implemented

PR: #30229 — fix(provider-cursor): stream live shell output

Two details from the plan above changed while writing it against the replay harness:

shellOutputText decodes the oneof directly and returns "" for any case other than stdout/stderr. The flat-key fallback is gone: it never matched a real frame, and keeping it would hide the next SDK shape change instead of failing the fixture.
The assertion did not need parameterizing. The Cursor transcript streams line 1 too, so the existing final-vs-live endsWith check already requires every line, and the Codex/Claude expectations stay in one place. The change is the transcript plus one provider row.

The transcript includes an exit frame after the stdout chunks, so treating the exit status as text fails the test, and a \r-rewritten progress line so the live tail is checked after terminal normalization.

Verification

Before the fix (main @ cefb544af3):

$ vp test run apps/server/src/orchestration-v2/testkit/CursorReplayFixtures.integration.test.ts -t "command_output_streaming/cursor"
AssertionError: the running command must stream its output: expected undefined to not equal undefined

After:

$ vp test run apps/server/src/orchestration-v2/testkit/CursorReplayFixtures.integration.test.ts -t "command_output_streaming"
     Tests  1 passed | 11 skipped (12)

$ vp test run apps/server/src/orchestration-v2/testkit/CodexReplayFixtures.integration.test.ts -t "command_output_streaming"
     Tests  1 passed | 24 skipped (25)

$ vp test run apps/server/src/orchestration-v2/testkit/ClaudeReplayFixtures.integration.test.ts -t "command_output_streaming"
     Tests  1 passed | 37 skipped (38)

$ vp test run packages/provider-cursor/src/server/adapter.test.ts
     Tests  21 passed (21)

$ vp run check:changed
  pass  format / lint / knip / graphql / typecheck (10 packages) / test (2 files)

Hi maintainers, let me know if theres anything you'd like me to do here.
I'd love to work more on this issue and become a contributor to the t3 community as well, the pr is already done and good to go.

would love a vouch!

Navigation Menu

Repository navigation

[Bug]: Cursor shell commands never stream live output #30228

Before submitting

Area

Steps to reproduce

Expected behavior

Actual behavior

Impact

Version or commit

Environment

Logs or stack traces

Workaround

Proposed implementation plan

Cursor live shell output — Implementation Plan

Constraints and non-goals

Invariants the fix must preserve

Baseline (must fail before the fix)

Task 1 — Add the failing Cursor replay fixture

Task 2 — Read the Cursor `{ case, value }` payload

Task 3 — Verification

Risks

Rollback

Implemented

Verification

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[Bug]: Cursor shell commands never stream live output #30228

Description

Before submitting

Area

Steps to reproduce

Expected behavior

Actual behavior

Impact

Version or commit

Environment

Logs or stack traces

Workaround

Activity

Zburgers commented on Nov 3, 2026

Proposed implementation plan

Cursor live shell output — Implementation Plan

Constraints and non-goals

Invariants the fix must preserve

Baseline (must fail before the fix)

Task 1 — Add the failing Cursor replay fixture

Task 2 — Read the Cursor { case, value } payload

Task 3 — Verification

Risks

Rollback

Zburgers commented on Nov 3, 2026

Implemented

Verification

Zburgers commented on Nov 4, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Task 2 — Read the Cursor `{ case, value }` payload