菜单
Analyze the current agent conversation transcript to estimate the number of input, output, and cumulative tokens consumed, including caching estimates.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | estimate-conversation-tokens |
| description | Analyze the current agent conversation transcript to estimate the number of input, output, and cumulative tokens consumed, including caching estimates. |
Use this skill when you or the user wants to inspect the token consumption of the current conversation session. This skill runs a Dart script to analyze the conversation's transcript_full.jsonl file and provides a detailed breakdown of step-by-step token usage.
Run the Dart token estimation script from the project root using the current conversation ID.
The script dynamically picks up the ANTIGRAVITY_CONVERSATION_ID environment variable:
dart .agents/skills/estimate-conversation-tokens/scripts/estimate_tokens.dart
If you need to estimate tokens for a different or past conversation, you can pass the conversation ID as a command-line argument:
dart .agents/skills/estimate-conversation-tokens/scripts/estimate_tokens.dart <conversation_id>
Report the outputs directly to the user in a well-formatted markdown table and summary list.
Ponder architectural designs, explore potential solutions, weigh tradeoffs, or analyze the impact of proposed codebase changes without making code changes. Always outputs the analysis to a Markdown artifact.
Autonomous pull request review loop that pushes code, polls for AI/bot review comments (e.g., Gemini Code Assist), surgically remediates feedback, commits, pushes, comments `/gemini review`, and loops until zero feedback remains. Requires the `github-pr-triage` skill.
Introspects conversation history to suggest new tools, skills, MCP servers, or identify context window hogs and optimization opportunities.
Synthesizes conversation history and active tasks into a visual hierarchy map (`sidequest.md`) to prevent context drift and cognitive overload across long sessions. Supports multiple sequential and concurrent main quests, sub-quests, and side-quests. Use when the user invokes `/sidequest`, asks where we are, what we were doing, or what's on our stack, or when the conversation branches across multiple topics, blockers, or digressions. Don't use for simple one-off questions that don't involve multi-step work or task hierarchy management.
Intake multi-step tasks by clarifying ambiguities, summarizing understanding, and confirming readiness via ask_question before executing code changes.
Use gob-curl and Buildbucket tools to inspect the status, tryjobs, and CI results of a Gerrit CL.