📖 pi-atelier 实战指南 — 从零教会你使用 pi-atelier 扩展生态，包含完整示例和最佳实践。

English | 程序中文文档

pi-session-analyzer

源码仓库 | npm

Session search and analysis for pi — search historical sessions, reconstruct timelines, audit behavior, and generate takeover reports.

Why You Need It

pi generates rich session logs (JSONL), but they're scattered across directories and hard to navigate. When you need to:

• Find what you did last week
• Continue work from a previous session
• Check who modified a file and why
• Audit an agent session for rule violations

...you'd have to manually dig through JSONL files. pi-session-analyzer gives the AI structured access to every session pi has ever run.

How It Works

┌─── session_search ──────────────────────────────┐
│                                                  │
│  grep ──→ 全文搜索所有会话（支持正则）           │
│  file ──→ 查找修改过特定文件的会话               │
│  list ──→ 列出最近 N 个会话                      │
│                                                  │
└──────────────────────┬───────────────────────────┘
                       │
                       ▼ 找到 sessionId
┌─── session_analyze ─────────────────────────────┐
│                                                  │
│  summary ──→ 元信息 + 摘要（首次分析首选）       │
│  entries ──→ 条目列表（支持 grep + 分页）        │
│  timeline ─→ 时间线（自动标注 [B1][B2] 分支）    │
│  chain ────→ 子代理调用链追踪                    │
│  audit ────→ 规则违规检查                        │
│  digest ───→ user/assistant 对话序列             │
│  branches ─→ 并行分支分析（/tree 产生）          │
│  takeover ─→ 5 维接手报告                        │
│  raw ──────→ 原始 JSONL 数据                     │
│                                                  │
└──────────────────────────────────────────────────┘

Installation

pi install git:github.com/catlain/pi-session-analyzer

Prerequisite: pi must be installed.

Tools

session_search

Parameter	Type	Required	Default	Description
action	"grep" | "file" | "list"	✅	—	Search mode
query	string	grep/file	—	Keyword (grep) or file path (file)
limit	number	❌	20	Max results
editOnly	boolean	❌	false	grep mode: only search edit/write operations

Examples:

session_search(action: "grep", query: "roadmap")
session_search(action: "file", query: "src/index.ts")
session_search(action: "list", limit: 10)
session_search(action: "grep", query: "error", editOnly: true)

session_analyze

Parameter	Type	Required	Default	Description
sessionId	string	✅	—	Session ID (supports prefix match)
action	enum	✅	—	See action table below
limit	number	❌	20	Max entries
offset	number	❌	—	Start position (0-based, entries mode)
grep	string	❌	—	Keyword filter (entries mode, supports regex)

Action reference:

Action	Output	Best for
summary	Metadata + abstract	First look at a session
entries	Filtered entry list	Finding specific events
timeline	Chronological with branch labels	Understanding session flow
chain	Sub-agent call trace	Debugging sub-agent workflows
audit	Rule violations	Checking agent behavior
digest	User/assistant conversation	Reading what happened
branches	Per-branch analysis	Analyzing /tree forks
takeover	5-dimension handoff report	Continuing from a past session
raw	Raw JSONL entries	Low-level debugging

Examples:

session_analyze(sessionId: "abc123", action: "summary")
session_analyze(sessionId: "abc123", action: "entries", grep: "error|fail")
session_analyze(sessionId: "abc123", action: "takeover")
session_analyze(sessionId: "abc", action: "entries", offset: 0, limit: 10)  # pagination

Storage

Session data is stored in pi's data directory:

Data	Location
Session JSONL files	~/.local/share/pi-coding-agent/sessions/
Per-session entries	{sessionId}.jsonl
Session ID matching	Prefix match — pass abc to match abc123def...

Use Cases

Scenario	Workflow
"What did I do last week?"	list → summary → digest
"Who changed this file?"	file(query: "path/to/file") → timeline
"Continue where I left off"	takeover → get 5-dimension handoff (intent, files, steps, next steps, decisions)
"Why did this break?"	entries(grep: "error") → audit
"What did the sub-agent do?"	chain → trace sub-agent execution

Best Practices

✅ Recommended

• Always start with summary — it's fast and gives you the full picture before drilling in
• Use grep filter in entries mode to find specific events without loading everything
• Use prefix match for sessionId — no need to type the full ID
• For large sessions (>100 entries), use offset + limit for pagination instead of loading all at once
• Use takeover when switching machines or resuming work the next day

❌ Not Recommended

• Don't use raw for large sessions — use entries with pagination instead
• Don't confuse session_search actions (grep/file/list) with session_analyze actions — they're different tools
• Don't skip summary and go straight to raw — you'll waste context on irrelevant data

Limitations

Limitation	Detail
Read-only	Cannot modify session data, only analyze it
No real-time streaming	Analyzes completed sessions only
Prefix match ambiguity	Short prefixes may match multiple sessions
Audit rules are static	Custom audit rules require code changes

Architecture

pi-session-analyzer/
├── index.ts              # Entry: register session_search + session_analyze tools
├── core.ts               # Session resolution, JSONL reader, types
├── core-visible.ts       # Visible sub-agent file resolution
├── search.ts             # list + grep implementation
├── search-file.ts        # file search (find sessions editing specific files)
├── search-utils.ts       # Shared search helpers
├── analyze.ts            # summary/entries/timeline/chain/raw/branches
├── digest.ts             # User/assistant conversation extraction
├── audit.ts              # Rule violation checker
├── audit-rules.ts        # Audit rule definitions
├── audit-types.ts        # Audit-specific types
├── takeover.ts           # Takeover tool entry point
├── takeover-core.ts      # 5-dimension takeover report logic
├── tests/                # Unit tests
└── package.json

Dependencies:

• @pi-atelier/shared-utils (bundled) — truncatedResult for large outputs
• @earendil-works/pi-coding-agent — ExtensionAPI (peer)

License

MIT