smart-compact

name	smart-compact
description	Analyzes the session context and generates focused instructions for pi's /compact command so that important information survives compaction. Use when the context window is getting tight or when you want to tidy up the session.

Smart Compact

pi fork of the Claude Code smart-compact skill. Analyze the session context and generate /compact instructions aligned with the user's intent. pi's /compact [instructions] accepts optional instructions that focus the summary — this skill produces those instructions.

How pi compaction works (background)

• Manual: /compact [instructions] — the instructions focus the LLM-generated summary of older messages.
• Auto: triggers when context tokens exceed contextWindow - reserveTokens (default reserve: 16384). The most recent keepRecentTokens (default: 20000) are kept verbatim; only older messages are summarized. Auto-compaction always uses the default summary prompt — custom instructions only apply to manual /compact, which is why running this skill before the threshold hits keeps you in control of what is preserved.
• Settings: ~/.pi/agent/settings.json or <project>/.pi/settings.json, under the compaction key (enabled, reserveTokens, keepRecentTokens).
• Summary shape: pi produces a structured summary (Goal / Constraints & Preferences / Progress / Key Decisions / Next Steps / Critical Context, plus <read-files> / <modified-files> lists tracked cumulatively across compactions). Instructions steer what lands in those sections.

Execution Flow

Phase 1: Session context analysis

Analyze the whole conversation history and extract:

• Tasks in progress: the implementation/fix/investigation currently underway
• Important technical decisions: architecture choices, library choices, design judgments
• Unresolved issues: errors, bugs, and problems still needing attention
• File change history: files changed/created during the session and why
• Context-dependent information: assumptions and constraints the follow-up work needs

Show the analysis result as a concise bullet list.

Phase 2: Interview

Based on the Phase 1 analysis, use pi-harness's Claude-compatible AskUserQuestion tool with multiSelect: true to ask the user what to keep. If interactive UI is unavailable, ask in plain conversation with a numbered list and state that multiple selections are fine.

Rules for generating the choices:

• Generate the choices dynamically from the session content
• Make each choice concrete, containing this session's specific context (task names, file names, tech stack, etc.)
• Offer 2–4 choices, plus "other" free-form input

Question template:

Which information should compaction preserve with priority? (multiple selections OK)

Example choices (generated dynamically from the session):
1. Implementation progress and remaining work for <specific task>
2. Error-investigation context around <file name>
3. Design decisions and rationale for <technology>
4. API spec and type definitions for <feature>

Important: derive the choices from the session analysis — never use preset generic choices.

Phase 3: Instruction generation

Based on the user's answers, generate the instructions to pass to /compact.

Generation guidelines:

1. Include the user-selected items with top priority
2. Include the minimum context needed to continue the current task
3. Include concrete proper nouns: file names, function names, error messages
4. Phrase as "what to preserve", not "what to discard"
5. Write the instructions in Japanese (the user's working language)
6. Focus on older context: pi keeps the most recent ~keepRecentTokens verbatim anyway, so spend the instructions on earlier decisions and findings that would otherwise be summarized away

Output format:

以下の情報を重点的に保持してコンテキストを圧縮してください：

1. [保持項目1の具体的な内容]
2. [保持項目2の具体的な内容]
3. [保持項目3の具体的な内容]

特に以下は正確に保持してください：
- [重要な固有名詞、パス、コマンドなど]

Phase 4: Hand off to the user

Present the generated instructions to the user:

Copy and run the following command:

/compact <generated instructions>

Usage Example

> /smart-compact

=== Phase 1: Session context analysis ===

Main contents of this session:
- Implementing the MCPMerger feature of the dotfiles CLI
- Fixing a deep-merge bug involving defu
- Adding tests in tests/core/mcp-merger.test.ts

=== Phase 2: Interview ===

Which information should compaction preserve with priority? (multiple selections OK)

1. Investigation history and fix for the MCPMerger defu-integration bug
2. Added test cases in mcp-merger.test.ts and their intent
3. claude.json merge spec (key-preservation rules)
4. Other

> user answers: 1 and 3

=== Phase 3: Instruction generation ===

Generated instructions:
---
以下の情報を重点的に保持してコンテキストを圧縮してください：

1. MCPMergerでdefuを使ったmcpServersマージ時に、既存キーが上書きされるバグの原因と修正内容（src/core/mcp-merger.ts の mergeConfig 関数）
2. claude.jsonのマージ仕様：mcpServersキーのみマージし、apiKey等の他キーはdefu経由で保持する設計判断

特に以下は正確に保持してください：
- src/core/mcp-merger.ts, tests/core/mcp-merger.test.ts のファイルパス
- defu の非破壊マージ挙動（ターゲット側の値が優先される仕様）
---


Copy and run the following command:

/compact 以下の情報を重点的に保持してコンテキストを圧縮してください：...

Notes

• /compact is a pi built-in command that only the user can run — same limitation as in Claude Code: the skill generates the instructions, and the user copies and executes the command manually
• Most effective when the session is long, i.e. context is approaching contextWindow - reserveTokens; run it before auto-compaction fires, since auto-compaction cannot take custom instructions
• To buy time before auto-compaction triggers, reserveTokens / keepRecentTokens can be tuned in the settings files listed above