Gemini Delegate (ACP)

Description

--- name: gemini-delegate description: Delegate coding tasks to Gemini CLI via ACP (Agent Client Protocol) for prototyping, debugging, code review, and web search. Structured JSON output with tool calls, thoughts, and plans. Permission-controlled file access. Multi-turn sessions. Proactive safety net before expensive operations. metadata: short-description: Delegate to Gemini CLI (ACP) --- # Gemini Delegate (ACP) Use Gemini CLI as a collaborator via the Agent Client Protocol. Claude owns the final result and must verify changes locally. ## Prerequisites ```bash pip install --user -r ~/.agents/skills/gemini-delegate/requirements.txt ``` Requires Gemini CLI >= 0.36.0 with `--acp` support. ## Core rules - Gemini is a collaborator; you own the final result. - Always use the bridge script (`scripts/gemini_bridge.py`). - **Use `--prompt-stdin`** for all automated calls to prevent deadlock and temp-file collisions. - **Use `--output-file AUTO`** when result persistence is needed (produces unique `0600` file). - Capture `SESSION_ID` from output and reuse it for follow-ups. - Default timeout: set Bash tool `timeout_ms` to **600000 (10 minutes)**. ## Quick start ```bash echo "Review src/auth.py around login() and propose fixes." | \ python3 ~/.agents/skills/gemini-delegate/scripts/gemini_bridge.py \ --cd "." --prompt-stdin --output-file AUTO ``` Output: JSON with `success`, `SESSION_ID`, `agent_messages`, `tool_calls`, `stop_reason`, `token_estimate`. ## Pre-delegation checklist (mandatory) Before EVERY delegation: 1. **Input size**: >200 lines → send file paths, not content. >500 lines → split the task. 2. **Staleness**: `git diff --stat <file>`. Unchanged + prior result exists → reuse, don't re-delegate. With `--cache`: auto-invalidates on commits and dirty tree; omit during active editing. 3. **Scope output**: Every prompt MUST include an output format (JSON preferred), a length constraint, and at least one `DO NOT` clause. 4. **Select model**: Always pass `--model` explicitly (see routing table below). ## Task routing | Task | Template | Model | Est. tokens | |---|---|---|---| | Diff review (<100 lines changed) | Focused diff review | `gemini-3-flash-preview` | 500-2k | | Code review (single file, <200 lines) | Review | `gemini-3-flash-preview` | 1k-3k | | Multi-file / architectural review | Tool-assisted review | `gemini-3.1-pro-preview` | 2k-8k | | Architecture analysis | Analysis / Plan | `gemini-3.1-pro-preview` | 3k-10k | | Bug diagnosis (>3 files) | Tool-assisted review | `gemini-3.1-pro-preview` | 2k-8k | | Web search / current info | Web search | `gemini-3-flash-preview` | 1k-3k | | Generate a patch (single file) | Patch | `gemini-3-flash-preview` | 1k-4k | | Adversarial plan critique | Plan review | `gemini-3.1-pro-preview` | 2k-5k | | Novel algorithm design | Analysis / Plan | `gemini-3.1-pro-preview` | 3k-10k | | File reads, grep, enumeration | Analysis / Plan | `gemini-3-flash-preview` | 500-2k | ## When to delegate ALL must hold: task reads >200 lines Claude hasn't seen, task is self-contained, output is verifiable. **High-value**: web searches, large file analysis, parallel reviews, adversarial second opinions, tedious enumeration, **plan review gate participation** (see below). **Proactive use in plan review gates**: When `/plan-review-gate` runs, include a Gemini reviewer as an additional adversarial reviewer alongside the standard reviewers — from Round 1, not as a later addition. Use `gemini-3.1-pro-preview` with the "Adversarial plan critique" template. ## When NOT to delegate ANY disqualifies: <30s of Claude reasoning, needs conversation context, file <50 lines, need to edit immediately after, security-sensitive, already done this session. ## Feedback log After verifying output, log the outcome via `--log-feedback "VERDICT|TASK_TYPE|EST_TOKENS|NOTE"`. Before delegating, scan last 10 entries for this task type — if rejection rate >50%, escalate model tier. Log immediately after verification; do not batch or skip. See `assets/reference.md` for format details. ## CLI flags | Flag | Description | Default | |---|---|---| | `--prompt` | Prompt text | | | `--prompt-file` | Read prompt from file | | | `--prompt-stdin` | Read prompt from stdin (mandatory for automation) | | | `--session-id` | Resume a specific session | | | `--new-session` | Force fresh session | | | `--sessions-dir` | Override session storage directory | | | `--cd` | Workspace root directory | `.` | | `--timeout` | Total max wall-clock seconds | 600 | | `--idle-timeout` | Max seconds between chunks | 120 | | `--first-chunk-timeout` | Max seconds for first chunk | 300 | | `--verbose` | Print heartbeat markers to stderr | | | `--output-file` | Write JSON to file (`AUTO` for unique temp) | | | `--approve-edits` | Allow Gemini to write files within `--cd` scope | | | `--cache` | Enable result caching. Opt-in. | off | | `--cache-ttl` | Cache TTL in seconds (1-2592000) | 86400 | | `--clear-cache` | Clear cache and exit | | | `--parallel-models` | Comma-separated models for parallel runs | | | `--log-feedback` | Append feedback entry (VERDICT\|TASK_TYPE\|EST_TOKENS\|NOTE) | | | `--model` | Gemini model to use | | ## Output format ```json { "success": true, "SESSION_ID": "abc-123", "agent_messages": "Gemini's response text", "tool_calls": [...], "stop_reason": "end_turn", "token_estimate": { "input_tokens": 1250, "output_tokens": 340, "model": "gemini-3-flash-preview", "estimated_cost_usd": 0.000391, "note": "Estimate only. Actual billing may differ." } } ``` ## Permission control By default, Gemini can **read** files within `--cd` scope but **cannot write**. `--approve-edits` enables scoped writes. Terminal execution is always blocked. ## References - `assets/prompt-template.md` — prompt templates for each task type - `assets/reference.md` — sessions, concurrency, timeouts, feedback log format, parallel review, collaboration capsule

Security Status