session-analyzer

name	session-analyzer
skill_id	session-analyzer
skill_version	0.3.0
version	0.3.0
category	reporting
description	Debug and teach agentic coding: a deterministic-first session timeline + cost report, with optional narrative polish and a standalone JSX visualiser.
when_to_use	when analyzing a completed Claude session to understand cost breakdown, agent behavior, tool usage timeline, or to generate a visual session report
updated_at	2026-06-12T00:00:00.000Z
tags	["session","reporting","cost","teaching","debugging"]
effort	medium
progressive_disclosure	{"entry_point":{"summary":"Reconstruct a Claude session's full timeline and cost breakdown deterministically from JSONL transcripts, with optional LLM prose polish and a standalone JSX visualiser.","when_to_use":"When analyzing a completed Claude session to understand cost breakdown, agent behavior, tool usage timeline, or to generate a visual session report.","quick_start":"1. Locate the session JSONL transcript 2. Run the deterministic extract 3. Review the timeline + cost report 4. (Optional) polish prose 5. (Optional) render the JSX visualiser"},"references":["references/schema.md"]}

Session Analyzer

Produce a deterministic-first session report for debugging and teaching agentic coding with Claude MPM. The report reconstructs an entire session's timeline — every PM turn, subagent call, skill use, MCP call, the model behind each call, tokens in/out (plus cache), and a TOTAL estimated cost at public rack rates — directly from Claude Code's JSONL transcripts.

The core data is generated with zero LLM inference. Inference is an optional, opt-in polish step that only rewrites human-readable prose; the machine-readable data is never touched by a model.

When to use this skill

• Post-session debugging — understand what the PM actually did: which subagents ran, what they reported back, where time and tokens went, which MCP calls were made and how they responded.
• Teaching effective MPM usage — turn a real session into a readable, annotated timeline that shows good (and bad) delegation, skill use, and cost patterns.
• Cost analysis — see per-model and per-call token/cost breakdowns and a grand-total estimate at rack rates, split into PM vs. subagent cost.

The hybrid workflow

The workflow has three stages. Stages 1 and 3 are fully deterministic. Stage 2 is the only place inference is allowed, and it is entirely optional.

1. Extract (deterministic, no inference)

Run the offline reporter to parse the JSONL transcripts and emit the canonical Markdown report. This stage alone produces a complete, valid, usable report — token, cost, model, agent, skill, and MCP data, all with zero LLM inference.

# Most-recent session for the current project directory
claude-mpm session-report

# A specific session
claude-mpm session-report --session <UUID>

# Custom output path (use - for stdout)
claude-mpm session-report --session <UUID> --output /tmp/report.md

# A different project directory
claude-mpm session-report --project /path/to/project

Equivalent module form: uv run python -m claude_mpm.cli session-report ...

Flags:

• --session <id> — session UUID; defaults to the most recent session for the project.
• --project <path> — project directory whose transcripts to read; defaults to the current directory.
• -o, --output <file.md> — output path; - writes to stdout.