| name | gemini-router |
| description | Delegate a task to Gemini CLI (`gemini -p`). Preset for long-context reading, research, translation review, and adversarial second-opinion passes. The 1M+ context window is its key affordance. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["routing","gemini","research","translation","long-context"],"related_skills":["claude-code-router","codex-router"]}} |
gemini-router
When to use: tasks whose input is huge (entire codebase audit, full paper PDF, long transcripts) OR tasks that need an independent third opinion distinct from Anthropic + OpenAI model families (per Regulations § Triple-Heterogeneous Review).
Inputs
task — prompt
model — leave unset to use Gemini CLI's own default. Explicit gemini-3.1-pro returns 404 from this account; the CLI auto-resolves to a working alias when -m is omitted.
working_dir — optional; if set, Gemini's tool calls run from there
effort — low / medium / high, defaults to medium
timeout_seconds — defaults to 1500
Invocation recipe
The Gemini CLI has no --effort flag (verified gemini --help 0.39.1). Use --approval-mode plan to keep the run read-only when auditing, otherwise omit. Always </dev/null for cron-safety.
TASK="${1:?task required}"
WORKING_DIR="${2:-$PWD}"
TIMEOUT="${TIMEOUT_SECONDS:-1500}"
APPROVAL="${APPROVAL_MODE:-default}"
cd "$WORKING_DIR" 2>/dev/null
GEMINI_M_ARG=""
[ -n "${MODEL:-}" ] && GEMINI_M_ARG="--model $MODEL"
timeout "$TIMEOUT" gemini -p "$TASK" $GEMINI_M_ARG \
--approval-mode "$APPROVAL" </dev/null
Output contract
VERDICT: <pass|fail|flag|ambiguous> on the final line
- Optional
EVIDENCE: <one-paragraph> for review tasks
When to choose this over the others
| Need | Pick |
|---|
| Cross-check a translation for fluency | gemini-router |
| Read 200k-token codebase to find a pattern | gemini-router |
| Adversarial third pass on a Codex-reviewed PR | gemini-router |
| Author code | claude-code-router |
| Run + fix tests | codex-router |
Anti-patterns
- ❌ Do NOT use Gemini for code authoring — its tool-use loop is weaker than Claude Code / Codex on this workload
- ❌ Do NOT remove
</dev/null