Slimeyd Gemini Delegator Gemini Delegator

Gemini Delegator

You orchestrate. Route work to Gemini whenever a task is independent, well-defined, and doesn't need conversation context — Gemini gets faster turnaround and parallel throughput, and you preserve your own context for synthesis and judgment. This is the default posture, not something to do only when asked.

Should I delegate this?

Is this task...
  ├─ Independent + well-defined (clear input, clear success)? → DELEGATE
  ├─ Writing — code, components, copy, prototypes?            → DELEGATE
  ├─ Extraction — pull JSON from text/docs?                   → DELEGATE to Gemma 4 (unlimited TPM)
  ├─ Research with citations?                                 → DELEGATE to 2.5 Flash + --enable-tools
  ├─ Batch — many items, same operation?                      → DELEGATE in parallel
  ├─ Multi-perspective review / council-of-experts?           → DELEGATE
  └─ Needs your orchestrator's tools or live credentials?     → KEEP

How to delegate

python3 -m gemini_key_pool.gemini_agent \
  --task "Your detailed prompt here" \
  --output /tmp/result.md

Picking a model — quick rules

• Default code/writing workhorse: gemini-3.5-flash — 20 RPD per project, good thinking depth.
• Batch + structured extraction: gemma-4-31b (unlimited TPM, 1500 RPD). Cheap to stuff giant prompts in.
• Research with citations: gemini-2.5-flash + --enable-tools (Search grounding is paid on Gemini 3.x — 2.5 is the free path).
• High volume / small tasks: gemini-3.1-flash-lite — 500 RPD is the best free quota.
• Paid models (Pro / image / video / agentic): not callable from a free-tier key. Use a paid Google Cloud key if you need them.
• Full quota numbers and 429 diagnosis: references/quotas.md.

Useful flags

Stack-aware delegation

Before writing the task prompt, think about what context Gemini will need — Gemini is stateless and has no access to your project, conventions, or domain-specific rules.

1. Identify which existing rule files apply. If you have project-level instructions for your orchestrator (e.g. CLAUDE.md, GEMINI.md, AGENTS.md), gather the relevant excerpts.

2. Identify which existing skills encode the relevant patterns. For example, React/Next.js skills, Postgres skills, brand-voice skills. If you have skill SKILL.md files locally, include them.

3. Pass everything via repeated --context-file (or --context-glob). The agent aggregates them into a single context block with ## Context: <path> headers:

   python3 -m gemini_key_pool.gemini_agent \
     --task-file /tmp/my-task.md \
     --context-file path/to/CLAUDE.md \
     --context-file path/to/relevant-skill/SKILL.md \
     --context-file path/to/code-file.tsx \
     --output /tmp/out.md
   

4. Write a sharp task prompt that points to the context and names the desired output.

Don't curate "the relevant subset" of skill content. Gemini's TPM is generous within a single request (250K on flash models, unlimited on Gemma). Dump the relevant context in full; spend effort on making the task prompt sharp instead.

Shell escaping when prompts get long

Inlining a multi-paragraph --task string into bash is fragile (quote breakage, backtick interpolation, dollar-sign expansion). Prefer:

python3 -m gemini_key_pool.gemini_agent \
  --task-file /tmp/my-task.md \
  --context-file /tmp/ctx.md \
  --output /tmp/out.md

If you're piping or composing on the fly, this also works:

python3 -m gemini_key_pool.gemini_agent \
  --task "$(cat /tmp/my-task.md)" --context-file /tmp/ctx.md --output /tmp/out.md

Crafting delegation prompts

Gemini agents have no conversation context — everything they need must be in the prompt or context file. Delegation-specific things to remember:

• Self-contained. No references to "the file I just read" or "as we discussed." Gemini hasn't read or discussed anything.
• Name the output format explicitly. Your orchestrator is going to parse the result, so spell out the structure: fenced code blocks with // FILE: markers, strict JSON shape, markdown with H2/H3 sections, etc.
• State anti-patterns + project quirks explicitly. Gemini doesn't know them. "Use shadcn/ui primitives in src/components/ui/, do not roll a new component" beats "follow project conventions."
• Watch shell escapes. Use --task-file for anything over a few lines.

Sensitive paths are filtered

--context-file and --context-glob paths are filtered against a built-in deny-list of substrings (.env, .ssh, .aws, credentials, secrets, id_rsa, id_ed25519, private_key, service-account, etc.). A matching path is refused with a 🛡️ log line. This is a guardrail against accidentally sending a .env to Gemini, not a substitute for thinking about what you pass in.

Validating Gemini output

Non-negotiable. Gemini agents are fast but can hallucinate — especially on:

• File paths and directory structures (may invent paths)
• Feature descriptions (may describe features the product doesn't have)
• Code details (verify against actual source)
• CLI flags (may invent plausible-looking argparse flags that don't exist — grep the source to confirm before relying on them)

Always read the output file and cross-check key claims before presenting to the user. For code suggestions, verify they compile / run. For factual claims, spot-check against source material.

Parallel dispatch

Launch multiple agents in parallel by running them as background processes. Throughput ceiling = (number of GCP projects in your pool) × (per-model RPM). With a single project's 8 keys, peak is the per-model RPM (5 or 15 depending on the model); with N projects, multiply by N.

Orchestration patterns

The basics:

• Parallel-Aggregate: dispatch 3–5 agents on the same input from different angles (security / performance / quality), synthesize their reports yourself.
• Council of Experts: a specialized Parallel-Aggregate where each agent adopts a domain-expert persona (UX, SEO, accessibility, etc.).
• Sequential Pipeline: stages where each step's output feeds the next (research → analyze → write).
• Batch Processing: same operation across many items; use Gemma 4 for the unlimited TPM.

For advanced patterns (Hierarchical, Double-Diamond, Handoff Chain, Structured Debate), see references/patterns.md.

Key audit (June 19, 2026 deadline)

Unrestricted keys stop working June 19, 2026. Every key in .env must be restricted to "Generative Language API" in AI Studio before then. The package ships an audit tool:

python3 -m gemini_key_pool.audit_keys

Run weekly. Writes a report to logs/key-audit-YYYY-MM-DD.md with per-project status and a checklist for keys that need action.

What Gemini agents cannot do

• Access your conversation history (everything must be in the prompt)
• Interact with the user mid-task
• Maintain state between separate calls
• Use your orchestrator's tools (Read, Edit, etc.) — pass file contents via --context-file
• Install custom Python packages (only ~25 pre-installed: numpy, pandas, scikit-learn, scipy, opencv, tensorflow, geopandas, matplotlib, ...)
• Code execution sandbox has a 30-second timeout

References

Source: SlimeyD/gemini-delegator — distributed by TomeVault.