# cursor-review — a review pass driven by the Cursor CLI (spec §9.7).
#
# Copy into {config_dir}/workflows/ for every project, or into
# .vincent/workflows/ inside one repo (project scope shadows global, §8.1).
#
# Two things differ from a claude or codex workflow, and both are cursor
# properties rather than vincent's:
#
#   * There is no `effort:` field anywhere below. Cursor has no effort flag —
#     reasoning depth is part of the model id (`claude-sonnet-5-thinking-high`,
#     `gpt-5.4-mini-xhigh`). An `effort:` on a cursor step is ignored, and a
#     claude/codex effort value is rejected outright by the §8.2 catalog check.
#     Run `cursor-agent models` to see what your account offers.
#
#   * `model: auto` is spelled out rather than left unset. Cursor persists the
#     model it is given to ~/.cursor/cli-config.json, so vincent always sends
#     one; naming it here keeps the workflow's behavior readable instead of
#     depending on an adapter default.
#
# Permission mode is left at the default (full-auto). `restricted` requires
# cursor's sandbox, which needs macOS or Linux — on Windows a restricted
# cursor step fails with `restricted_unsupported` rather than quietly running
# unrestricted (§9.4).
name: cursor-review
description: Review the working tree with Cursor, write the findings to a file, and commit them.

defaults:
  agent: cursor
  model: auto
  max_retries: 1

steps:
  - id: review
    type: agent
    prompt: |
      Review this repository for correctness bugs and obvious simplifications.

      Task: {{.Task.Title}}
      {{if .Task.Description}}Context: {{.Task.Description}}{{end}}

      Write your findings to REVIEW.md in the repository root, most severe
      first, each with a file:line reference. If you find nothing worth
      reporting, write "No findings." and stop. Do not change any other file.

  - id: commit
    type: command
    # Single-quoted: the commit message contains ": ", which a plain YAML
    # scalar cannot carry.
    #
    # This step is also the verification. If the agent wrote nothing, `git
    # add` finds no REVIEW.md and the step fails with nonzero_exit — the task
    # blocks instead of reporting a review that never happened. Nothing more
    # portable is needed: `&&` behaves the same under bash, cmd, and
    # PowerShell 7, which is what §8.3 may hand a command step.
    run: 'git add REVIEW.md && git commit -m "review: {{.Task.Title}}"'

  - id: read
    type: manual
    instructions: |
      Read REVIEW.md on branch {{.Task.BranchName}} before merging anything.
      The agent ran full-auto: everything it did is in the task transcript.