token-saver
/token-saver
Ogranicza koszty tokenów w Antigravity z Gemini na GCP: porządkuje przestrzeń roboczą i pilnuje rozrastających się kontekstów.
- Autor
- swap-cameroon ↗
- Repozytorium
- kuro-antigravity-config ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 11.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 12
- Skrypty
- 4
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 11.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 4 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: evals/run_evals.py, resources/intercept_wasteful_prompts.py, test_interceptor.py, test_token_saver.py
Co robi i kiedy
Ogranicza koszty tokenów w Antigravity z Gemini na GCP: porządkuje przestrzeń roboczą i pilnuje rozrastających się kontekstów.
Oryginalny opis (EN)
Clever strategies and guidelines to optimize token consumption and reduce costs in Antigravity when using Gemini on GCP/Agent Platform. Activate this skill when the user requests cost control, token optimization, workspace pruning, or when context windows are ballooning.
Jak działa
- 1. Architectural & Tool Optimization
- 2. Multi-Model Routing
- 3. Vertex AI Context Caching
- 4. Workspace Hygiene & File Filtering
- 5. Session Hygiene & History Management
- 6. "Caveman Mode" — Output Compression
- 7. Dynamic Tool & Prompt Filtering
- 8. AGENTS.md Budget Rule
- 9. Automated Pre-Invocation Prompt Interceptor (Hook)
Podgląd SKILL.md
---
name: token-saver
description: |
Clever strategies and guidelines to optimize token consumption and reduce costs in Antigravity when using Gemini on GCP/Agent Platform.
Activate this skill when the user requests cost control, token optimization, workspace pruning, or when context windows are ballooning.
---
# Token-Saver: Cost & Token Optimization Guide
This skill provides actionable guidelines and templates for optimizing Gemini token usage in Antigravity, leveraging GCP Vertex AI context caching, model selection, and agent architecture best practices.
## 1. Architectural & Tool Optimization
* **Lean Tool Registration**: Do not enable unused MCP servers. Every active MCP server registers its tool definitions in the system prompt, adding to the static "token tax" on every single turn.
* **Minimal Tool Permissions**: When spawning subagents or running tasks, request only the narrowest possible permission scopes to avoid loading excessive resource metadata.
* **Context Slicing**:
- **Avoid reading full files**: Use the line-range arguments (`StartLine` and `EndLine`) in file-viewing tools to load only relevant functions or sections.
- **Limit Command Outputs**: When running commands, filter or truncate verbose output (e.g., use `git diff --stat` or limit log length via `git log -n 5`) to prevent piping thousands of lines of terminal output back into the prompt context.
## 2. Multi-Model Routing
* **Flash vs Pro**:
- Use `gemini-3.5-flash` or `gemini-3.5-flash-lite` for routine tasks (e.g., viewing files, formatting layouts, simple syntax fixes, listing directories, writing unit tests).
- Reserve `gemini-3.5-pro` (or heavier reasoning models) for complex planning, system design, or multi-file debugging.
* **Use `/btw` for Side-Queries**: In the Antigravity CLI, use the `/btw` command to ask quick questions. This prevents starting a full agent planning cycle, which executes multiple expensive reasoning turns.
## 3. Vertex AI Context Caching
* **Static Context Ordering**: To trigger GCP Vertex AI context caching, place large static contexts (such as library documentations or large system prompts) at the *very beginning* of the context payload.
* **Cache Lifetime**: Since context caching on GCP charges a lower rate for cached tokens, keep your session consistent and avoid changing initial system configurations or workspace paths mid-conversation, as doing so invalidates the cache.
## 4. Workspace Hygiene & File Filtering
* **Implement `.antigravityignore`**: Prevent the agent from indexing and parsing binary, generated, or dependency files. Create an `.antigravityignore` file in your workspace root.
* See the template in: [resources/example-antigravityignore](resources/example-antigravityignore)
## 5. Session Hygiene & History Management
* **Prune Conversation History**: Conversation history grows cumulatively with each turn. Use the `/clear` or `/new` commands to clear the chat context once a sub-task is completed.
* **Forking & Branching**: Use `/fork` or `/branch` to spin off a new conversation thread, carrying over only the current status instead of the full historical reasoning path.
* **Reuse Subagents**: Instead of launching a new subagent for every task, send a new message to an existing idle subagent via `send_message` to save startup overhead.
## 6. "Caveman Mode" — Output Compression
Inspired by [JuliusBrussee/caveman](https://github.com/JuliusBrussee/caveman). Compress what the agent *says*, not its reasoning quality.
**Rules when activated:**
* Strip pleasantries, apologies, filler ("Certainly!", "I'd be happy to...", "Let me explain...").
* Drop articles (*the*, *a*, *an*) and connective prose where clarity is preserved.
* Use symbols: `→` for causality/flow, `=` for equivalence/assignment.
* **Never compress**: code blocks, file paths, error messages, CLI commands, or safety warnings.
* **Auto-clarity override**: If a high-risk operation is detected (file deletion, security rule changes, production deployments), automatically revert to full verbose explanations.
**Realistic expectations:** Output token savings of ~8–20% in agentic workflows. Most agent output is code and tool calls which cannot be compressed.
## 7. Dynamic Tool & Prompt Filtering
* **Conditional Prompt Sections**: Only include instruction blocks relevant to the active tools. Don't load web development guidance for backend-only tasks.
* **Symbol-Level Navigation**: Use `grep_search` to locate specific functions/classes *before* reading files. Avoid loading entire files to "find" something.
* **Session-Level Caching**: If the same tool call (e.g., `list_dir` on a stable directory) has already been made in this session, reference the prior result instead of re-executing.
* **Prune MCP Tool Schemas**: Each unused MCP tool adds ~200–500 tokens of schema to every turn. Disable MCP servers you are not actively using via `/mcp` or `mcp_config.json`.
## 8. AGENTS.md Budget Rule
* Keep your `AGENTS.md` file under **600 tokens**. It is loaded into *every* conversation.
* See the active workspace rules as a real-world example of a lean, optimized guideline prompt: [resources/example-AGENTS.md](resources/example-AGENTS.md)
* If it exceeds 2,000 tokens, refactor task-specific instructions into dedicated `SKILL.md` files that load on-demand.
* Periodically audit and prune `AGENTS.md` — AI-generated instructions tend to bloat over time.
## 9. Automated Pre-Invocation Prompt Interceptor (Hook)
* **Prompt Interception**: A registered global `PreInvocation` hook (`intercept_wasteful_prompts.py`) scans user prompts on every turn.
* **Vague/Wasteful Query Detection**: It checks for extremely vague prompts (e.g., short words like "go", "do it", "test" with no details) or massive print/dump requests (e.g., "explain every line in the codebase", "print everything") or infinite loops.
* **Active Safeguard**: When a wasteful query is intercepted, the hook injects a system warning instructing the agent to **halt expensive tool execution** and prompt the user for clarification, preventing hundreds of dollars in wasted token expenses.
Pliki w paczce (12)
- README.md 8 KB
- SKILL.md 6 KB
- benchmark_report.md 5 KB
- test_interceptor.py 3 KB
- test_token_saver.py 30 KB
- evals/benchmark_thresholds.json 1 KB
- evals/interceptor_evalset.jsonl 4 KB
- evals/run_evals.py 8 KB
- evals/results/eval_results.json 9 KB
- resources/example-AGENTS.md 2 KB
- resources/example-antigravityignore 1 KB
- resources/intercept_wasteful_prompts.py 4 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · swap-cameroon/kuro-antigravity-config | 100 | MIT |
| ikdeaplop/public-generalized-products | 100 | MIT |
| haiderejaz6692/codex-powerpack | 100 | MIT |
| cyberspezie/Token-Saver | 100 | MIT |
| ladansojudi/token-saver | 100 | Apache-2.0 |
| punyawatt51/token-saver | 90 | MIT |
| kreht1/anchor | 90 | MIT |
| lokikill123/codex-token-skills | 90 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 664d6e4c2702 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100token-saver-config
Kompresuje wyjście poleceń CLI w oknie kontekstu, ograniczając liczbę tokenów o 60 do 99 proc. - 100token-saver-config
Konfiguruje ustawienia kompresji tokenów, sprawdza status procesora i diagnozuje hooki oraz statystyki oszczędności. - 100handoff
Zapisuje .token-saver/handoff.md, żeby praca mogła wrócić w świeżej, taniej sesji, gdy kontekst się kończy. - 100working-with-gemini
Opisuje Gemini CLI i Antigravity CLI: rozszerzenia, skille, hooki, plik GEMINI.md, modele i konfigurację. - 100token-cost-analyzer
Analizuje i optymalizuje koszty API LLM: zużycie tokenów, audyt kosztów i prompty pod kątem wydatków na AI. - 100antigravity
Przydziela zadania modelowi Gemini w Antigravity przez lokalny most plików, sprawdza jego sygnał życia i samodzielnie odbiera wynik.