hermes - 💡(How to fix) Fix [Feature]: claude-code-print and cursor-agent-print SKILL.md for non-ACP coding CLI delegation [1 participants]

Problem or Use Case

Hermes ships first-class skills for copilot, claude-code, codex, and opencode — but only copilot works as a true ACP child via delegate_task(acp_command="copilot"). The other three coding CLIs (claude, cursor-agent, codex) do not implement the ACP wire protocol:

$ claude --help | grep -E "acp|stdio"
# (no matches — claude only supports --print/-p)

$ cursor-agent --help | grep -E "acp|stdio"
# (no matches — cursor-agent only supports --print/-p)

This means Hermes users who want to spawn a Claude or Cursor child agent (e.g. to use a Claude Max subscription, or burn Cursor quota for parallel work) are stuck with two awkward paths:

1. terminal() one-shot — fire-and-forget, no streaming, no structured output parsing
2. process() long-lived — manual stdin/stdout marshaling, no result extraction

Neither integrates with the existing delegate_task orchestration surface (sub-session tracking, tool-call accounting, parent context isolation), and there is no SKILL.md teaching the agent the right invocation patterns. PR #13303 (copilot-cli skill) closes the documentation gap for Copilot but the same gap exists for claude -p and cursor-agent -p.

Related work that doesn't cover this:

• #2653 — adds acp_command override to delegate_task (already implemented). Solves spawning ACP-speaking children, but claude/cursor-agent are not ACP-speaking.
• #5258 (DRAFT) — generalizes CopilotACPClient for 14 ACP-compatible CLIs. Same scope limitation.
• #13303 — copilot-cli SKILL.md. Excellent template, but Copilot-only.

Proposed Solution

Two optional-skills/ modules following the PR #13303 template, no Python changes required:

optional-skills/autonomous-ai-agents/claude-code-print/SKILL.md

Documents three invocation modes for the claude CLI:

1. One-shot print mode (claude -p "<prompt>" --output-format json) — for short tasks Hermes can fully describe upfront. Works through terminal() today.
2. Streaming print mode (claude -p ... --output-format stream-json --include-partial-messages) — for tasks where intermediate progress matters. Works through process() today.
3. Background interactive (claude in tmux) — for multi-turn tasks where Hermes wants to inject follow-up messages.

Plus the relevant flags Hermes should always set: --allowedTools, --add-dir <cwd>, --model, --effort, --max-budget-usd, --json-schema for structured output, and a comparison matrix vs copilot / codex / opencode.

optional-skills/autonomous-ai-agents/cursor-agent-print/SKILL.md

Same shape for cursor-agent, covering --mode plan|edit|chat, --sandbox, --worktree-base, --approve-mcps, --trust, --workspace, the --output-format text|json selector, and --list-models for model discovery.

Both skills go into optional-skills/ (per CONTRIBUTING.md guidance) since they require third-party CLIs the user must install themselves — not bundled, but discoverable via hermes skills browse.

Alternatives Considered

1. Extend agent/copilot_acp_client.py to detect non-ACP CLIs and switch from JSON-RPC to a simple stdin/stdout pipe. I prototyped this locally (~115 lines, adds a _COMMAND_SIMPLE_PIPE_ARGS registry mapping claude → -p --output-format json, cursor-agent → -p --output-format text, codex → exec). It works but conflicts with the architectural direction in #5258 (one client per ACP CLI, generalized) and per CONTRIBUTING.md ("Should it be a Skill or a Tool?", almost always skill). Skill route is cleaner.

2. Build a generic "PrintModeClient" sibling to ACPClient that implements the same delegate_task interface for non-ACP CLIs. Significant new Python surface — better as a follow-up only if the skill route proves insufficient.

3. Wait for upstream CLIs to add ACP — claude and cursor-agent show no signs of it, and ACP support is an opinionated choice (see Anthropic's stance on first-party clients vs third-party automation).

The skill route is the smallest change, ships immediately, and matches the pattern established by #13303.

Feature Type

New bundled skill

Scope

Small (single file, < 50 lines) — actually two SKILL.md files at ~250 lines each, but no Python.

Contribution

• I'd like to implement this myself and submit a PR

I'm happy to author both SKILL.md files following the structure of skills/autonomous-ai-agents/copilot-cli/SKILL.md from PR #13303. Will pre-flight-check claude/cursor-agent flag references against the latest CLI releases as PR #13303 did.

Debug Report

Not applicable — this is a missing-capability request, not a runtime issue.