Cursor Delegate Setup Guide¶

Repo: matdev83/hermes-cursor-dispatcher Version: v1.1.0 Stars: 1 License: MIT Created: July 19, 2026 Install: npx skills add matdev83/hermes-cursor-dispatcher --skill cursor-delegate -g -y

What It Is¶

A skill for Hermes Agent that safely delegates repository-level coding, debugging, refactoring, or implementation work to Cursor CLI (agent). The skill:

Transfers exact UTF-8 prompts through a Python wrapper (no shell interpretation)
Prefers isolated Git worktrees for implementation
Captures structured results with bounded log tails
Requires Hermes to independently inspect diffs, run tests, and accept/reject
Integrates natively with Grok models (default: grok-4.5-xhigh)

Key design rule: Cursor is an external implementation executor - never a source of truth. Hermes owns task construction, workspace selection, safe invocation, diff review, independent tests, corrective iterations, and the final acceptance decision.

Prerequisites¶

Cursor CLI installed and in $PATH (agent --version must work)
Git 2.30+
Python 3.10+ (for the wrapper script)
Hermes Agent with terminal toolset access

Verify Cursor CLI:

agent --help
agent --version

Installation¶

npx skills add matdev83/hermes-cursor-dispatcher --skill cursor-delegate -g -y

This places the skill at $HERMES_HOME/skills/cursor-delegate/.

Verify:

ls $HERMES_HOME/skills/cursor-delegate/
# Should show: SKILL.md  scripts/  templates/

How It Works¶

Wrapper Script¶

The bundled wrapper is $HERMES_HOME/skills/cursor-delegate/scripts/cursor_delegate.py. It:

Reads an exact UTF-8 prompt from a file
Sends it to Cursor CLI through stdin (no shell interpretation)
Uses shell=False, explicit cwd, a narrowly allowlisted environment
Runs with a timeout in an isolated process group
Captures bounded log tails (mode-0600 prefixes)

Why Not Shell Interpolation?¶

Never invoke Cursor by interpolating arbitrary prompt text into a shell command. The wrapper avoids: - Per-argument size limits on Linux - Shell metacharacter injection - eval, shell=True, or echo as prompt transport - Environment variable prompt smuggling

Cursor CLI Contract¶

The wrapper assumes Cursor CLI supports: - Executable: agent - Headless mode: --print - Output formats: text, json, stream-json - Read-only modes: --mode plan, --mode ask - Edit mode: default (no --mode edit flag needed) - Workspace selection: --workspace <path> - Headless trust: --trust - Model selection: --model <id> - Prompt input through stdin

Default model: grok-4.5-xhigh (verified against installed CLI).

The wrapper accepts --mode edit but deliberately omits it from Cursor's argv for edit mode. It does not pass --force, --yolo, --approve-mcps, --add-dir, or arbitrary headers.

Mandatory Workflow¶

Step 1: Inspect Before Delegating¶

Hermes must first inspect: - Repository structure and applicable instruction files - Relevant modules, interfaces, and implementation constraints - Current tests and repository-specific validation commands - Current branch, git status --short, repository root, and HEAD

Step 2: Workspace Selection¶

Choose the appropriate workspace: - Detached worktree (preferred): Isolates Cursor's changes from the main working tree - Current working directory: Only if the repo is clean and the change is trivial

# Preferred: isolated worktree
git worktree add --detach /tmp/cursor-work-$(date +%s) HEAD

Step 3: Write Prompt to File¶

Write the exact prompt to a temporary file (never inline in shell):

prompt = """Task: Refactor the database connection pool in src/db/pool.py
Context: Current implementation has a race condition under high concurrency.
Requirements:
1. Use connection-per-thread pattern
2. Add timeout on acquisition
3. Maintain backward compatibility with existing callers
4. Run existing tests: pytest tests/db/ -x
"""
write_file("/tmp/cursor_prompt.txt", prompt)

Step 4: Invoke the Wrapper¶

python3 $HERMES_HOME/skills/cursor-delegate/scripts/cursor_delegate.py \
  --workspace /tmp/cursor-work-20260719 \
  --prompt-file /tmp/cursor_prompt.txt \
  --timeout 600 \
  --model grok-4.5-xhigh

The wrapper returns structured JSON:

{
  "exit_code": 0,
  "stdout_tail": "...",
  "stderr_tail": "...",
  "timeout": false,
  "log_file": "/tmp/cursor_delegate_20260719_120000.log"
}

Step 5: Diff Review¶

Hermes MUST independently inspect the diff:

cd /tmp/cursor-work-20260719
git diff
git diff --stat

Do NOT accept Cursor's report without inspecting actual changes.

Step 6: Independent Tests¶

cd /tmp/cursor-work-20260719
pytest tests/db/ -x -v

Step 7: Accept or Reject¶

Accept: Apply the diff to the main branch, commit with attribution
Reject: Discard the worktree, document why, possibly retry with adjusted prompt

Safety Boundaries¶

The wrapper enforces:

Boundary	Mechanism
No shell injection	Prompt from file, `shell=False`
Output bounds	50KB stdout, 50KB stderr hard caps
Process isolation	Separate process group, kill on timeout
No secret leakage	Prompt file is mode 0600, deleted after run
No privilege escalation	Never passes `--trust` without explicit user approval
Workspace isolation	Detached worktree, never on dirty repo

The wrapper does NOT pass: - --force - no forced overwrites - --yolo - no skip-confirmation - --approve-mcps - no MCP auto-approval - --add-dir - no directory injection - Arbitrary headers - no HTTP header manipulation

When to Use¶

Use for: - Non-trivial implementation, debugging, migration, or refactoring - Repository changes where a second coding executor adds value - Tasks with large Markdown prompts, code blocks, or structured data - Independent implementation attempts in isolated Git worktrees

Do NOT use for: - One-line local edits Hermes can safely make directly - Operations outside a clearly selected workspace - Prompts containing secret values (API keys, tokens, passwords) - Arbitrary MCP access or implicit privilege escalation - Accepting Cursor's report without inspecting actual changes

Error Recovery¶

Cursor CLI not found¶

which agent || echo "Cursor CLI not installed. Install from cursor.com/cli"

Timeout¶

The wrapper has a configurable timeout (default: 600s). If exceeded: 1. The Cursor process group is killed 2. Partial output is captured in the log file 3. The worktree may contain partial changes - discard and retry

Model not available¶

agent --list-models
# Verify grok-4.5-xhigh or override with --model

Worktree conflicts¶

# Clean up stale worktrees
git worktree list
git worktree remove /tmp/cursor-work-* --force

Pitfalls¶

Never trust Cursor's self-report - Always inspect the actual diff. Cursor may claim success while producing broken code.
Dirty repos - The wrapper does NOT check for uncommitted changes. Hermes must verify git status --short before delegating.
Model alias drift - grok-4.5-xhigh may not appear in agent --list-models but was verified working. If it fails, check model availability with agent --list-models and override with --model.
Large repos - Worktree creation on large repos (1GB+) may take several seconds. Account for this in timeout calculations.
Tests may not exist - If the repo has no tests for the changed area, Hermes must manually verify correctness through code review.

Verification Checklist¶

[ ] agent --version returns a version number
[ ] agent --help shows supported flags
[ ] Skill directory exists: ls $HERMES_HOME/skills/cursor-delegate/
[ ] Wrapper exists: ls $HERMES_HOME/skills/cursor-delegate/scripts/cursor_delegate.py
[ ] python3 $HERMES_HOME/skills/cursor-delegate/scripts/cursor_delegate.py --help shows usage
[ ] git worktree --help confirms git version supports worktrees

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects