context-mode/configs/antigravity-cli/rules/context-mode.md

context-mode routing for Antigravity CLI

context-mode MCP tools are installed for this agy session. Use them when the task analyzes, counts, filters, compares, searches, parses, transforms, fetches, or otherwise processes data. Raw bytes should stay in the sandbox; only the derived answer should enter the conversation.

Do not dump — derive (most common mistake)

Do NOT use context-mode/ctx_execute_file or ctx_execute to print a whole file or a full method/config (e.g. print(FILE_CONTENT), cat config.yaml, dumping a 281-line file). That defeats the purpose: on agy the tool's stdout is saved to a step file that you then read back, so a full dump costs your context window TWICE (once in the result, once on the follow-up read). Print only the specific value, matches, count, or known line-range you need:

• WRONG: ctx_execute_file(path:"config.yaml", language:"python", code:"print(FILE_CONTENT)")
• RIGHT (value): code:"import yaml; d=yaml.safe_load(FILE_CONTENT); print(d['active_strategy'])"
• RIGHT (matches): code:"import re;[print(i+1,l) for i,l in enumerate(FILE_CONTENT.splitlines()) if 'active_strategies' in l]"
• RIGHT (known slice): code:"print(chr(10).join(FILE_CONTENT.splitlines()[190:230]))"

If you truly need to read a small, exact byte range to edit it, native Read / view_file on that range is fine — but never dump an entire file through a hook.

Tool call surface

Antigravity CLI exposes context-mode tools as context-mode/<tool> calls. If the host uses the generic MCP wrapper, call call_mcp_tool with:

• ServerName: "context-mode"
• ToolName: "ctx_execute", "ctx_execute_file", "ctx_batch_execute", "ctx_fetch_and_index", "ctx_search", or "ctx_index"
• Arguments: a JSON object for that tool

Do not read ~/.gemini/antigravity-cli/mcp/context-mode/*.json to discover schemas. Those files are agy's cached tool schemas and reading them spends context. Use these argument shapes instead:

• context-mode/ctx_execute: {"language":"python","code":"..."}
• context-mode/ctx_execute_file: {"path":"path/to/file","language":"python","code":"..."}
• context-mode/ctx_batch_execute: {"commands":[{"label":"...","command":"..."}],"queries":["..."]}
• context-mode/ctx_fetch_and_index: {"url":"https://...","source":"..."}
• context-mode/ctx_search: {"queries":["q1","q2"]}
• context-mode/ctx_index: {"path":"path/to/file-or-dir","source":"..."} or {"content":"...","source":"..."}

Mandatory routing

• Think in code. For analyze/count/filter/compare/search/parse/transform tasks, write code with context-mode/ctx_execute and print only the final answer. Program the analysis; do not read raw data and compute mentally.
• File read for analysis: there is no separate ctx_read tool. context-mode/ctx_execute_file is the context-mode file-read surface. It loads the file into FILE_CONTENT inside the sandbox. Print only selected lines, counts, matches, summaries, or structured results. Never print FILE_CONTENT wholesale unless the user explicitly asks for a full file dump.
• Native Read / view_file is correct when editing requires exact bytes, or when a small known range is needed. For analysis, exploration, summarization, counting, filtering, or searching inside a file, use context-mode/ctx_execute_file.
• Use one context-mode/ctx_batch_execute call for multi-command repository reconnaissance. One batch should replace many shell/list/search calls.
• Use context-mode/ctx_execute for shell commands whose output may exceed a short fixed answer. Native Bash is only for git, mkdir, rm, mv, navigation, installs, or short observable output.
• Use context-mode/ctx_fetch_and_index for web content, then context-mode/ctx_search to query it. Do not dump raw HTML into the conversation.
• Use context-mode/ctx_index only when content should be stored and searched later. For a one-off file question, prefer context-mode/ctx_execute_file; for follow-up retrieval, index with a descriptive source and query with context-mode/ctx_search.
• Return only derived answers, concise summaries, selected snippets, or file paths to written artifacts. Do not paste raw command dumps, full files, large search results, cached schemas, or raw HTML into the conversation.