Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Development

whats-next

Re-orient the user after they return to a session — or after they open a FRESH context pointed at a worktree/branch — and let them decide the next step by answering lightweight structured questions (Yes/No, A/B/C/D via AskUserQuestion) instead of facing a blank prompt. Trigger when the user invokes /whats-next (optionally naming a worktree, branch, or repo path), or asks "接下来做什么", "下一步是什么", "我该做什么", "这个 worktree/分支做到哪了、下一步呢", "what's next", "what should we do next", "帮我决定下一步", or returns after a long gap and wants to resume work with minimal typing. Works with zero conversation history: state is then rebuilt from git evidence (diffs, branch commits, uncommitted changes) and handoff/task docs. NOT for pure recap with no decision needed (that is a recap request — answer it directly or use a context-rebuild skill if available), and NOT for questions about a specific file or piece of code.

lycfyi · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add lycfyi/yskills --skill whats-next

About this skill


name: whats-next description: >- Re-orient the user after they return to a session — or after they open a FRESH context pointed at a worktree/branch — and let them decide the next step by answering lightweight structured questions (Yes/No, A/B/C/D via AskUserQuestion) instead of facing a blank prompt. Trigger when the user invokes /whats-next (optionally naming a worktree, branch, or repo path), or asks "接下来做什么", "下一步是什么", "我该做什么", "这个 worktree/分支做到哪了、下一步呢", "what's next", "what should we do next", "帮我决定下一步", or returns after a long gap and wants to resume work with minimal typing. Works with zero conversation history: state is then rebuilt from git evidence (diffs, branch commits, uncommitted changes) and handoff/task docs. NOT for pure recap with no decision needed (that is a recap request — answer it directly or use a context-rebuild skill if available), and NOT for questions about a specific file or piece of code. license: MIT

What's Next

The user runs many sessions in parallel and comes back after hours away. They cannot remember where this conversation stands. The job: rebuild the state FOR them, then hand them a small set of concrete, clickable decisions — never a blank prompt to fill in.

Match the language of the conversation.

Workflow

0. Pick the evidence source

Two invocation shapes, same downstream flow:

  • Warm session — there is real conversation history: the conversation IS the record. Go straight to step 1.
  • Cold start / named target — a fresh context, or the user points at a worktree, branch, or repo path ("这个 worktree 的 diffs", "新分支做到哪了"). There is no conversation to mine; rebuild state from repo evidence instead:
    • Resolve the target first. If ambiguous, git worktree list / branch list. Don't infer the base/target branch from branch names alone — check the remote HEAD or whatever tool of record the setup has. One clarifying question max.
    • Read, in rough priority order: uncommitted changes (git status + diff of the working tree), branch commits not on the base branch (git log base..HEAD with --stat), the actual diff vs. base when commit messages aren't enough, any task/handoff doc that scopes this work (a handoff doc in the repo, the PR description, an issue-tracker ticket named in commits), and recent-commit trailers that record acceptance/deploy state.
    • The mapping: committed-and-merged → DONE; committed on branch but not merged → done-but-not-landed; uncommitted diff → in flight; task-doc items with no corresponding diff → planned, not started. A step described in a doc with no code evidence did NOT happen.
    • Don't read every file end-to-end — read enough to name the milestones and the fork points, and say when a judgment rests on commit messages alone.

1. Reconstruct state (silently)

From the evidence source chosen above, determine:

  • The overall goal of this session, in one sentence.
  • What is already DONE and verified (committed? deployed? approved?). A big code or rule change that is deployed but not yet manually verified by the user is NOT done — classify it as awaiting-manual-verification, a state of its own.
  • What is IN FLIGHT: background agents, workflows, deploys, external waits. Never fabricate a result for anything still running — report it as pending.
  • What is BLOCKED and on what — especially things blocked on the user's own decision.
  • What was PLANNED but not started.

Trust the conversation record over memory of intent: if a step was described but no tool call shows it happened, it did not happen.

2. Brief recap first (3–5 lines, no more)

Before asking anything, print a tight orientation so the choices make sense:

  • First line, always: the close-session verdict. State plainly whether this session can be closed right now — "✅ 可以关:没有在跑的任务,成果都已落盘" or "⚠️ 先别关:XX 还在跑 / XX 只存在于这个对话里还没写进文件"。Judge it by:
    • Anything still running (background agents, workflows, deploys being watched)?
    • Any work product that exists ONLY in this conversation — not yet written to a file, committed, or recorded anywhere durable?
    • Any pending step that needs THIS session's context to finish, vs. one a fresh session could pick up from files/handoff docs alone? If closing loses nothing, say so explicitly — permission to close is the answer the user most often needs. If not closable, name the single cheapest action that would make it closable (e.g. "把结论写进 handoff 文档就能关").
  • One line: what this session is about.
  • One line: last completed milestone.
  • One line each: anything in flight or blocked.
  • Standard SOP — post-deploy manual test. If the state includes a bigger code or rule change that was deployed but not yet manually verified, attach a quick manual test recipe right in the recap: 3–5 steps — which page/entry point (with URL), what account or data to use, what to do, what the expected result looks like, and what result would mean it's broken. One of the AskUserQuestion options should then be "walk through this manual test" (usually the recommended one). Deployed-but-unverified also weighs into the close-session verdict — it's the classic "don't close yet" reason.

Plain prose, outcome-first, no headers, no internal codenames or table names. The recap exists only so the user can answer the questions confidently — details they don't need for the decision stay out.

3. Identify the REAL fork points

List the decisions that actually determine what happens next. Good fork points:

  • "Deploy now vs. run acceptance first"
  • "Continue feature X vs. switch to the bug that surfaced"
  • "Commit what's staged vs. review the diff together"
  • "The blocked item: resolve it via A or drop it"

Not fork points (never ask these):

  • Anything with an obvious conventional answer — just do it.
  • Anything already decided earlier in the conversation — don't re-litigate.
  • Fake choices where every option means "
Read the complete source ↗