feat(skills): add gemini-cli delegation skill

[spideystreet]

What does this PR do?

Adds a bundled delegation skill for Google Gemini CLI, mirroring the existing pattern of skills/autonomous-ai-agents/claude-code and skills/autonomous-ai-agents/codex.

This lets Hermes delegate coding work to the gemini CLI the same way it already delegates to Claude Code and Codex — one-shot tasks, background long-runs, parallel worktree fixes, and PR reviews.

Distinct from PR #11270: that PR added Gemini as an inference provider (Gemini-as-a-model). This PR is orthogonal — it delegates to the gemini CLI as an autonomous agent, not using Gemini as Hermes' brain.

What Gemini brings that the existing delegation skills don't:

• 1M-token context (whole-repo audits in one pass)
• Native --approval-mode plan for stronger read-only guarantees than allowlists
• No git-repo requirement (Codex needs one)
• No PTY required for -p mode (cleaner than Codex's tmux/PTY plumbing)

Related Issue

Fixes #10036

Type of Change

• 🎯 New skill (bundled or hub)

Changes Made

• Add skills/autonomous-ai-agents/gemini-cli/SKILL.md (173 lines)
  • Frontmatter follows the project standard (name, description, version, author, license, metadata.hermes.tags, metadata.hermes.related_skills)
  • Documents: prerequisites, one-shot tasks, four approval modes, read-only audits with 1M context, JSON/stream-json output, background mode, worktree-based parallel work, PR-review patterns, batch reviews, model selection, subcommands, rules

How to Test

Tested end-to-end against gemini CLI v0.38.1 on macOS (Darwin 25.1.0) before opening this PR. Each documented pattern was verified individually.

# 1. Confirm Hermes discovers the skill
hermes skills list | grep gemini-cli
# Expected: gemini-cli | autonomous-ai-agents | local | local

# 2. Inspect frontmatter
hermes skills inspect autonomous-ai-agents/gemini-cli

# 3. End-to-end delegation
mkdir -p /tmp/gem-test && echo "TODO: refactor" > /tmp/gem-test/file.py
hermes chat -Q -s autonomous-ai-agents/gemini-cli --max-turns 4 \
  -q "Use the gemini-cli skill to audit /tmp/gem-test for TODO comments. Use plan mode. Report in 1 line."
# Expected: Hermes invokes `gemini -p ... --approval-mode plan` and reports the TODO

Verified working patterns from the SKILL:

• ✅ gemini -p one-shot (no PTY, no git repo required)
• ✅ --approval-mode plan (read-only enforced — no writes possible)
• ✅ --approval-mode yolo (auto-approves writes; created a real file)
• ✅ -o json / -o stream-json (parseable envelope with token + cost stats)
• ✅ --include-directories (multi-dir workspace)
• ✅ Background via shell & (no PTY needed for -p)
• ✅ git worktree add + per-worktree gemini (parallel issue fixing)
• ✅ Piped stdin (git diff | gemini -p ...)
• ✅ gemini mcp list, extensions list, skills list subcommands

Caveat documented in SKILL: built-in -w/--worktree flag is interactive-only (silently dumps help when combined with -p). The skill documents git worktree add as the scripted alternative.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (feat(skills): ...)
• I searched for existing PRs — no existing PR adds a gemini-cli delegation skill (PR feat(gemini): add Google Gemini CLI OAuth inference provider #11270 is the inference-provider integration, different scope)
• My PR contains only changes related to this skill (no unrelated commits)
• I've tested on my platform: macOS 15 (Darwin 25.1.0), Hermes Agent v0.10.0, gemini CLI v0.38.1
• N/A — pytest tests/ -q: this PR is a bundled skill (Markdown only), no Python code paths changed
• N/A — Tests for changes: skill is documentation; end-to-end test steps included in "How to Test"

Documentation & Housekeeping

• N/A — No cli-config.yaml.example change needed (skill auto-discovered)
• N/A — No CONTRIBUTING.md / AGENTS.md change needed
• I've considered cross-platform impact: skill uses standard CLI flags; gemini CLI is npm-installable on macOS/Linux/Windows, brew on macOS/Linux. Background & syntax is bash/zsh; the rule "no PTY needed for -p" applies cross-platform
• N/A — No tool descriptions/schemas changed

For New Skills

• This skill is broadly useful — delegation to Gemini CLI is in the same category as claude-code and codex, both already bundled. Free-tier auth (60 req/min, 1k req/day) means most users can try it without paid keys
• SKILL.md follows the standard format (frontmatter, prerequisites, patterns/sections, rules)
• No external dependencies that aren't already available — only requires gemini CLI (npm or brew install) and the existing Hermes terminal + process tools
• I've tested the skill end-to-end with hermes chat -s autonomous-ai-agents/gemini-cli -q "..." — working

Screenshots / Logs

End-to-end run on macOS, Hermes v0.10.0, model xiaomi/mimo-v2-pro:

$ hermes chat -Q -s autonomous-ai-agents/gemini-cli --max-turns 4 \
    -q "Use the gemini-cli skill to audit /tmp/gem-e2e for TODO comments. Use plan mode. Report in 1 line."
session_id: 20260418_202739_fe5b63
1 TODO found: file.py:1 — "TODO: refactor this".

[teknium1]

Thanks for adding the missing Gemini CLI delegation path. Current main still has no skills/autonomous-ai-agents/gemini-cli/ skill, so the request in #10036 remains valid.

Problems

• skills/autonomous-ai-agents/gemini-cli/SKILL.md:3 has a 180-character, multi-sentence description; AGENTS.md:888-900 requires one sentence of at most 60 characters.
• SKILL.md:1-10 omits platforms: and credits only Hermes Agent. New skills require audited platform gating and external contributor-first credit (AGENTS.md:916-931). This matters because the examples include POSIX-oriented /tmp and shell-background usage.
• The document omits the required modern section order and no tests/skills/test_gemini_cli_skill.py accompanies it (AGENTS.md:933-950).

Suggested changes

• Modernize the frontmatter and section structure, add a focused skill test, and then salvage the otherwise isolated new-file change.

Automated hermes-sweeper review.

[teknium1]

Please reduce this to a single sentence of at most 60 characters. New and modernized skills must meet the hardline description limit in AGENTS.md:888-900.

[teknium1]

For an external contribution, credit the human contributor first (for example, spideystreet + Hermes Agent), per AGENTS.md:926-931.

[teknium1]

Add audited platforms: metadata and metadata.hermes.category. The document includes POSIX-oriented /tmp and shell-background examples, so the supported-platform declaration needs to reflect that audit.