Sequential thinking

Sequential Thinking

Record your reasoning as an explicit, numbered chain of thoughts in a state file, one CLI call per thought. This replaces the sequential-thinking MCP server: same call → ack → next-thought rhythm, but the chain lives on disk, so it survives context compaction and can be replayed later.

The script is at scripts/think.py inside this skill's directory. The working directory at runtime is the user's project, so always invoke it by absolute path (you know this skill's location from where you read this file):

python3 <this-skill-dir>/scripts/think.py <command> ...

Workflow

1. Start a chain — one line stating the problem, plus a step estimate:

   think.py new "为什么 X 指标在 3 月下跌" --total 5
   

2. Record every thinking step as you produce it. Do the thinking in the argument itself — the thought text should carry real analysis, not a label:

   think.py add "候选原因有 A/B/C，其中 B 与时间线吻合……"
   

3. Revise instead of forcing. When an earlier step turns out wrong, don't bend later reasoning around it — mark the correction explicitly:

   think.py add "第 2 步的假设不成立，因为……" --revises 2
   

4. Branch to explore alternatives when two solution paths both look viable:

   think.py add "换个思路：如果从需求侧看……" --branch-from 3 --branch-id demand-side
   

5. Adjust the estimate freely with --total N on any add — it is an estimate, not a budget. Add more thoughts after the "last" one if needed.

6. Finish only when a hypothesis has been generated and verified against the chain, and you are satisfied with the answer:

   think.py add "结论：……，已对照第 1/3/5 步验证" --done
   

7. Replay at any time with think.py show (or show --json). Use this to recover the full reasoning chain after context compaction, or to review the chain before writing the final answer.

Principles

• One thought per call; keep the ack rhythm. Each call returns {thoughtNumber, nextThoughtNeeded, ...} — while nextThoughtNeeded is true, keep thinking.
• Question earlier decisions; express uncertainty when it exists.
• Filter out irrelevant information rather than carrying it forward.
• Generate a solution hypothesis, then verify it against the chain of thought before --done. If verification fails, keep adding thoughts.

State file

Resolution order — pick the most session-scoped option available:

1. Preferred: if your harness gives you a session-specific scratchpad or temp directory, pass --file <scratchpad>/thoughts.json on every call. This isolates concurrent sessions that share a working directory, and the chain still survives context compaction within the session.
2. Fallback: with no --file, state goes to ${XDG_STATE_HOME:-~/.local/state}/think/<cwd-hash>/thoughts.json — one chain per working directory, nothing written into the user's project. Caveat: two concurrent sessions in the same directory would interleave into (and garble) one chain, so prefer option 1 whenever you can.

Rules either way:

• Always begin a task with think.py new — never add onto a chain you did not start. If an add ack contains a warning field (chain idle for hours), treat it as a stale leftover: run new.
• Thoughts are written to disk in plain text and outlive the session; do not put secrets (keys, credentials, private data) into thought text.
• new clears the previous chain; reset deletes the file; where prints the path.