Session Historian

Look up what happened in prior Claude Code sessions. Index and search raw JSONL transcripts. Cache session summaries.

Script

.claude/skills/session-historian/scripts/session_query.py

Run directly (shebang + PEP 723, no manual uv run prefix needed):

.claude/skills/session-historian/scripts/session_query.py list
.claude/skills/session-historian/scripts/session_query.py messages <session-id>
.claude/skills/session-historian/scripts/session_query.py search "experiment worktree"
.claude/skills/session-historian/scripts/session_query.py show <session-id>
.claude/skills/session-historian/scripts/session_query.py index

Workflow: "I forgot what happened"

1. List recent sessions to find the relevant one:

   .claude/skills/session-historian/scripts/session_query.py list --limit 10

2. Read verbatim user messages (fastest way to reconstruct context):

   .claude/skills/session-historian/scripts/session_query.py messages <session-id-prefix>

3. Search for specific topics across all raw files:

   .claude/skills/session-historian/scripts/session_query.py search "experiment" --limit 10

4. Generate and cache a session summary (when detail is needed):
   • Run show <session-id> to get session metadata and file path
   • Read the raw JSONL file
   • Produce a structured summary following the template below
   • Write to ~/.claude/kaizen/session-summaries/<session-id>.md
   • Run mark-summarized <session-id> to flag in the index

Summary Template

When generating a session summary, write to ~/.claude/kaizen/session-summaries/<session-id>.md:

---
source_type: claude-session
source_path: <file_path>
session_id: <session_id>
project: <project_name>
date_range: <first_ts> → <last_ts>
user_message_count: <N>
assistant_turns: <N>
confidence: high|medium|low
generated_at: <ISO timestamp>
---

## Summary

[BLUF: 2-3 sentences on what this session accomplished]

## User Messages (verbatim)

[Numbered list of every user message, exactly as written, with timestamp]

1. [2026-02-21T15:40] "First ensure that .claude/worktrees is gitignored..."
2. [2026-02-21T16:51] "Ah. Do you not remember the experiment..."

## Actions Taken

[Bullet list of what the AI did — file edits, commands run, commits made]

- Created .claude/worktrees directory and added to .gitignore
- Spawned sub-agent a0263e5 to read workflow docs → produced project_workflow.draft.md
- ...

## Sub-Agent Tasks and Outcomes

| Agent ID | Task | Outcome |
|----------|------|---------|
| a0263e5  | Read docs, produce workflow draft | Produced project_workflow.draft.md |
| a5d97a8  | Execute work-backlog-item workflow | Partial — blocked at AskUserQuestion node |

## Key Findings / Decisions

[Important decisions made, things discovered, conclusions reached]

## What Was NOT Completed

[Work started but not finished, blocked items]

## Sources

File: <file_path>
Indexed: <DB_PATH>

Fidelity Rules (from summarizer skill)

• Read before summarizing: Read the actual JSONL file, do not guess from metadata.
• Verbatim user messages: Copy exactly — never paraphrase.
• Preserve counts: "3 sub-agents spawned, 2 succeeded, 1 blocked" not "most sub-agents succeeded."
• Distinguish absence: "Not mentioned in transcript" not "didn't happen."
• Search raw files: search command reads raw JSONL, not summaries. Never search summaries for facts.

Index Location

~/.claude/kaizen/session-index.duckdb    ← DuckDB index (sessions + user_messages tables)
~/.claude/kaizen/session-summaries/      ← Cached AI-generated summaries

The index is built on first list or explicitly with index. It is NOT automatically kept current — re-run index or pass --rebuild to pick up new sessions.

Command Reference

errors

List tool errors from a session with timestamps and tool names.

Performs a two-pass scan: first builds a map of tool-use IDs to tool names from assistant records, then extracts tool_result blocks where is_error is True from user records and resolves the tool name for each.

# Rich formatted output (default)
.claude/skills/session-historian/scripts/session_query.py errors
.claude/skills/session-historian/scripts/session_query.py errors <session-id-prefix>

# Tab-separated plain text: timestamp TAB tool_name TAB error_content
.claude/skills/session-historian/scripts/session_query.py errors --raw

Arguments and options:

• session_id — session ID prefix or "last" (default) for the most recent session
• --raw — output tab-separated lines: timestamp\ttool_name\terror_content_single_line

Exit codes: 1 if session not found; 0 if session found but no errors exist.

tools

List all tools used in a session with call counts and success/failure breakdown.

Performs a two-pass scan: collects tool_use blocks from assistant records, then correlates tool_result blocks from user records to classify each call. Calls with no matching result are counted as unmatched (session interrupted before the tool returned).

# Rich table output (default) — sorted by total calls descending
.claude/skills/session-historian/scripts/session_query.py tools
.claude/skills/session-historian/scripts/session_query.py tools <session-id-prefix>

# Tab-separated plain text with header row
.claude/skills/session-historian/scripts/session_query.py tools --raw

Arguments and options:

• session_id — session ID prefix or "last" (default) for the most recent session
• --raw — outputs header tool\ttotal\tsuccesses\tfailures\tunmatched then one row per tool

Table columns: Tool Name, Total Calls, Successes, Failures, Unmatched.

Exit codes: 1 if session not found; 0 if session found but no tool calls recorded.

irritation

Detect user frustration signals in a session. Reports two signal types:

1. Correction phrases — user messages containing phrases such as "wrong", "stop", "undo", "revert", "no," etc., matched case-insensitively.
2. Stuck tool loops — sequences of 3 or more consecutive identical tool calls (same tool name and input hash) in assistant records.

# Rich formatted output — two sections: Correction Phrases + Stuck Tool Loops
.claude/skills/session-historian/scripts/session_query.py irritation
.claude/skills/session-historian/scripts/session_query.py irritation <session-id-prefix>

# Type-prefixed tab-separated lines
.claude/skills/session-historian/scripts/session_query.py irritation --raw

Arguments and options:

• session_id — session ID prefix or "last" (default) for the most recent session
• --raw — type-prefixed tab-separated lines:
  • Correction phrase match: phrase\t<ts>\t<phrase>\t<excerpt>
  • Stuck loop signal: loop\t<ts>\t<tool>\t<count>\t<hash>

Exit codes: 1 if session not found; 0 if session found but no signals detected.

current-path

Print the JSONL file path for the current live session. Reads CLAUDE_SESSION_ID from the environment and resolves the path using the encoded current working directory. Does not use DuckDB — pure filesystem check.

Default output is a single raw line (machine-readable), suitable for agent pipelines. Use --rich for human display.

# Raw output (default) — one absolute path, no markup, agent-safe
.claude/skills/session-historian/scripts/session_query.py current-path

# Rich Panel display
.claude/skills/session-historian/scripts/session_query.py current-path --rich

Arguments and options:

• No session_id argument — reads CLAUDE_SESSION_ID from the environment
• --rich — renders the path inside a Rich Panel with a green border (default: raw path only)

Exit codes: 1 if CLAUDE_SESSION_ID is not set, or if the expected JSONL file does not exist (prints diagnostics to stderr including session ID, expected path, and projects directory).

JSONL Schema Reference

Records in ~/.claude/projects/<project-slug>/<session-id>.jsonl:

{"type": "user", "sessionId": "uuid", "timestamp": "ISO8601",
 "message": {"content": "string or [{type,text}]"}}
{"type": "assistant", "sessionId": "uuid", "timestamp": "ISO8601",
 "message": {"content": [{...}]}}

Records with toolUseResult key are tool results — skip for user message extraction. Tool use blocks have type: "tool_use" inside message.content array.