Analyze Sessions

Tools for querying past pi sessions. All scripts are stdlib Python 3, no dependencies, and read directly from ~/.pi/agent/sessions/.

Data shape (one-liner)

Each session is a JSONL file. Records are session (header with cwd, id, timestamp), model_change, thinking_level_change, and message (roles: user, assistant, toolResult). Assistant messages carry usage.cost already split into input/output/cacheRead/cacheWrite/total — cost analysis is just summing those. Subagent transcripts live nested inside the parent session's directory.

Scripts

All scripts share the same filter vocabulary (see "Shared filters" below). Run them with python3 from anywhere:

python3 ~/.pi/agent/skills/analyze-sessions/scripts/<script>.py [args]

cost.py — cost rollups

Subagent costs are included by default so totals reflect actual spend. Pass --show-subagents to see the subagent share per row, or --no-subagents to exclude.

# Last 7 days, broken down by day (default)
python3 cost.py

# Last 30 days, top 10 projects by spend
python3 cost.py --since 30d --by project --limit 10

# Cost-per-model (each assistant message credited to its own model)
python3 cost.py --since 30d --by model

# The 10 most expensive sessions of the last month
python3 cost.py --since 30d --by session --limit 10

# Cost of one project, all time
python3 cost.py --cwd /path/to/your/project

# Grand total only
python3 cost.py --since 30d --by total

# Machine-readable
python3 cost.py --since 30d --by day --json

Groupings: total, day, project, model, session. When grouping, --limit caps groups, not sessions.

prompts.py — dump user prompts for pattern mining

Output is markdown grouped by project (--format jsonl available). Prompts above --max-chars are dropped because they're almost always pasted context, not actual prompting.

# Default: markdown dump, max 2000 chars per prompt
python3 prompts.py --since 30d

# Tighter cap, one prompt per JSONL line
python3 prompts.py --since 7d --max-chars 1500 --format jsonl

# One project's prompts
python3 prompts.py --cwd /path/to/your/project --since 30d

# Prompts that mention a topic
python3 prompts.py --grep "rate limit" --since 60d

The typical workflow for "find patterns I could turn into global instructions":

1. Run prompts.py --since 30d and read the output.
2. Group by recurring themes (same correction repeated across projects, same setup question, same complaint).
3. Propose additions to global CLAUDE.md / project AGENTS.md / pi instructions.

show_session.py — render one session as markdown

# The most recent session
python3 show_session.py --latest

# A specific session by id prefix (8 chars is enough)
python3 show_session.py --session 019e475b

# The most recent session in a project
python3 show_session.py --latest --cwd /path/to/your/project

# Include subagent transcripts inline below
python3 show_session.py --session 019e475b --include-subagents-content

# Drop thinking entirely / show fewer chars
python3 show_session.py --session 019e475b --max-thinking -1 --max-tool-output 800

Each tool result is fenced with …[N more chars elided]… if truncated. Default truncations: tool output 2000, assistant text 4000, thinking 600. Pass 0 to a limit to disable it, -1 to --max-thinking to omit thinking entirely.

search.py — search across transcripts

Substring by default, regex with --regex (smart-case). Searches both user and assistant text by default.

# Substring across everything
python3 search.py "supabase RLS"

# Only my prompts, last 60 days
python3 search.py "global instruction" --in user --since 60d

# Regex
python3 search.py --regex "TODO\\(.+\\)"

# More context per match
python3 search.py "rate limit" --context 2

Each hit prints the session header plus a python3 show_session.py --session <id> line so you can drill in directly.

Shared filters

Available on all four scripts:

Flag	Meaning
--since WHEN / --until WHEN	YYYY-MM-DD, ISO datetime, or relative: 7d, 2w, 3h, 30m
--cwd SUBSTR	Substring match on the session's real cwd. Repeatable.
--model SUBSTR	Substring match on model id. Repeatable.
--provider {anthropic,openai,google}
--session ID	Session id or prefix (8 chars usually unique)
--include-subagents / --no-subagents	Override the script default
--limit N	Cap items returned (caps groups, not sessions, for cost.py group views)
--min-cost USD	Drop sessions below this spend
--min-messages N	Drop short sessions
--errors-only	Only sessions with at least one toolResult.isError
--grep SUBSTR	Case-insensitive substring on the session's concatenated user prompts

Subagent defaults

• cost.py: included (totals = real spend)
• prompts.py, show_session.py, search.py: excluded (a subagent's "user" message is a task description written by another agent, not your prompt)

Common queries

Question	Command
Total cost in the last 7 days	python3 cost.py --since 7d --by total
Daily spend trend, last 30 days	python3 cost.py --since 30d --by day
Most expensive projects this month	python3 cost.py --since 30d --by project --limit 10
Most expensive sessions ever	python3 cost.py --by session --limit 10 --until 1d
Cost of one project	python3 cost.py --cwd /path/to/proj
Patterns in my prompting	python3 prompts.py --since 30d --max-chars 1500 → read the output
What did I do yesterday	python3 show_session.py --latest --since 1d
Where did the agent struggle	python3 cost.py --since 30d --errors-only --by session --limit 10
Find old session about X	python3 search.py "X"

Notes

• All paths are read-only; the scripts never modify session files.
• The library (scripts/sessions.py) is reusable: import it for ad-hoc analysis.
• A full scan over a few hundred sessions takes ~1–2 seconds. No caching.