daemon-backend-cursor v0.2.0A

Immutable. This exact content is served forever at /api/v1/blob/873a3e9c50064b03.

---
name: daemon-backend-cursor
description: >
  Nested daemon-cli-backends reference for the Cursor daemon backend's flag
  surface. Read this only when a daemon task needs Cursor-specific CLI flags
  (model selection, output/tooling switches): it routes you to the installed
  CLI's live help via shell and shows how to translate that help into the
  generic `backend_options` mechanism. It is not a flag catalog.
version: 0.2.0
last_changed_at: 2026-07-27T00:00:00Z
related_files:
- src/lingtai/tools/daemon/manual/reference/cli-backends/SKILL.md
- src/lingtai/tools/bash/manual/reference/bash-cursor-agent/SKILL.md
maintenance: |
  Tracks the Cursor daemon backend flag-discovery topic it documents; update when that integration changes.
---

# Cursor Daemon Backend — Flag Discovery Entrypoint

The installed CLI's own help is the authority for Cursor Agent flags; this page
is only the entrypoint. Conversion rules live in the parent
[`reference/cli-backends/SKILL.md`](../../../SKILL.md).

## Discover flags from the installed CLI

1. Load `shell-manual` (its nested `reference/bash-cursor-agent/SKILL.md` has
   broader Cursor Agent CLI context).
2. Run `agent --version` and `agent --help` in bash. The daemon spawns root
   `agent` directly — `agent -p --force --output-format stream-json <prompt>` —
   so root help is authoritative (keychain errors happen before help on some builds).
3. Translate what you found into `backend_options` with the parent's generic
   conversion rules (nothing Cursor-specific is added).

## Example: model selection via the generic route

`backend_options` keys become long flags, inserted after the harness-owned
flags and before the task prompt:

```jsonc
{
  "backend": "cursor",
  "tasks": [{
    "task": "Implement and validate the change.",
    "tools": [],
    "backend_options": {
      "model": "opus"
    }
  }]
}
// argv: --model opus
```

The flag and model vocabulary belong to the installed CLI — LingTai does
not validate, enumerate, or simulate them. Confirm `--model` and its
accepted values in your installed `agent --help` before relying on this; an
unknown flag or value is the CLI's error, not the daemon's.

## Source-pinned stream-json usage
For installed `agent-cli@2026.05.28-a70ca7c`, accept usage only from `type=result`, `subtype=success`, `is_error=false` with all four non-negative integer fields: `inputTokens`, `outputTokens`, `cacheReadTokens`, `cacheWriteTokens`.
`inputTokens` is already net; UI-only `cli_tokens` adds read + write as cached, preserves raw usage, ignores invalid/all-zero blocks, and counts duplicate terminal results once.
Join model only from a preceding matching `system/init` by `session_id`; provider stays unknown because this source emits no provider identity (do not infer it from `apiKeySource`, model, backend, credentials, or environment). This is version-pinned, not a cross-release claim.

## Subscription & auth

Cursor account/subscription (keychain login); LingTai does not inject
credentials — the CLI must be signed in.

Official docs: https://docs.cursor.com/agent

## Harness boundary

Cursor declares no reserved-flag list at validation, so nothing is refused
beyond generic key/value safety rules. Still, do not re-set harness-owned
surfaces: `-p` (non-interactive
print mode), `--force` (allows file modifications in print mode),
`--output-format stream-json` (one JSON event per stdout line — the daemon's
progress/result parser and the session-id capture depend on it), and
`--resume` (owned by `ask` follow-ups, which replay the captured session id).

Session and completion: the first session-id-shaped stream event is stored
as `cursor_session_id`; `ask` resumes with
`agent -p --force --resume <cursor_session_id> --output-format stream-json
<message>` (one follow-up in flight per session). Per-run MCP injection is
not wired for this backend yet, so no `finish` contract: success comes from
the stream's final result event and the process exit code.

`backend_options` is honored only at `emanate` time; `ask` follow-ups reuse the
session without re-passing it.