Skip to main content

goal

Helps you use the /goal persisted-goal workflow effectively. Use when you want the agent to pursue a multi-step objective autonomously across turns, with budget control and completion verification.

跳到安装

来源信息

仓库
ramarivera/pi-goal
最近来源活动
2026年6月9日 01:13
检测到的 SKILL.md 语言
英语
星标
5
分支
0

安装方式

默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。

检查来源文件

决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。

正在显示 SKILL.md

SKILL.md
来源说明 · 只读预览
name
goal
description
Helps you use the /goal persisted-goal workflow effectively. Use when you want the agent to pursue a multi-step objective autonomously across turns, with budget control and completion verification.
# Goal Use this skill when you want the agent to keep working toward an objective across multiple turns without you re-prompting after each step. ## When to use /goal Prefer the `/goal` workflow when: - The task spans multiple files, tests, or verification steps - You want the agent to self-continue after each turn until done - You need a concrete completion audit before the agent declares victory - You want budget control over long-running agentic work Do **not** use `/goal` for: - Single-turn questions or quick lookups (regular chat is faster) - Exploratory brainstorming where the end state is unclear - Tasks where you want to review every tool call before it runs ## Commands ```text /goal <objective> # Create an active goal /goal <objective> --budget 100000 # Create with token budget /goal status # Show current goal, usage, and budget /goal pause # Pause continuation (current turn finishes) /goal resume # Resume automatic continuation /goal clear # Clear the goal and stop continuation ``` ## Writing a good objective A good objective is specific enough that the agent can verify completion without guessing. **Weak:** "Fix the auth code" **Strong:** "Refactor all JWT verification in src/auth/ to use the new jose library. Update unit tests and integration tests. Run the test suite and confirm zero failures. Stop if any test fails." Include in your objective: - **Scope:** Which files, modules, or directories - **Success criteria:** What "done" looks like (tests pass, types check, linter clean) - **Constraints:** Budget limits, stop conditions, things to avoid - **Verification:** How to confirm the work is correct ## Model tools When the extension is loaded, agents can use these tools directly: - `get_goal` — read the current persisted goal state - `create_goal` — create a goal when none is active or paused - `resume_goal` — unpause or re-pressurize an active/paused goal and schedule continuation - `update_goal` — mark the current goal complete only after the completion audit passes ## The completion audit Before the agent marks a goal complete, it performs an audit: 1. Restate the objective as concrete deliverables 2. Map every requirement to actual evidence (files, test output, PR state) 3. Verify coverage — passing tests alone do not prove completion if the objective has requirements beyond test coverage 4. Identify any missing, incomplete, or unverified items 5. Only call `update_goal` with `status: "complete"` when the audit passes **You can help** by making your objective explicit about what evidence counts as done. ## Budget guidance Token budgets are the primary cost-control lever. Rough estimates: - **Small goal** (single module, ~20 files): 50,000–150,000 tokens - **Medium goal** (cross-module, ~50 files): 150,000–500,000 tokens - **Large goal** (codebase-wide, 100+ files): 500,000–2,000,000+ tokens Set your budget conservatively. You can always `/goal resume` to continue where you left off — goal state and progress survive across sessions. ## Lifecycle ``` User creates goal -> Agent works -> Turn ends -> Agent continues -> ... -> Audit passes -> Goal complete ^ | | | | v v v /goal pause Budget limit No-tool turn User input /goal resume -> stops -> suppressed -> resets suppression /goal clear gracefully ``` - **Paused** goals do not auto-continue - **Budget-limited** goals stop scheduling new turns gracefully (current turn wraps up) - **No-tool turns** suppress the next automatic continuation until you send input or resume - **User input** resets continuation suppression and lets the agent continue ## Recommendations - Prefer `/goal` over manual re-prompting for any task that takes more than 2–3 turns - Use `--budget` on your first `/goal` in a new codebase to learn the burn rate - Use `/goal status` mid-flight to check progress and remaining budget - Use `/goal pause` before switching to a side conversation or unrelated task - If the agent seems stuck in a loop, use `/goal pause`, assess, then `/goal resume` or `/goal clear` ## Examples ### Migration campaign ```text /goal Migrate all 47 React class components in src/components/ to functional components with hooks. Run the existing test suite after each file. Stop if any test fails. --budget 300000 ``` ### Test coverage gap ```text /goal Increase test coverage for src/payments/ from 43% to 80%. Write integration tests, not unit tests with mocked dependencies. Run coverage after each new test file. ``` ### Lint fix sweep ```text /goal Fix all 312 ESLint errors reported by `npm run lint`. Fix in batches of 10 files. Run lint after each batch to confirm the count decreases. ``` ### Pausing for a side question ```text /goal pause /btw what file defines this route? /goal resume ```
在 GitHub 查看