Session Analysis

Comprehensive analysis of Claude Code usage across the rhi ecosystem, generated January 2026 using normalize sessions stats --all-projects --by-repo and normalize sessions show --analyze.

2329 sessions | 119,146 tool calls | ~$4,177 estimated cost (Sonnet)

Aggregate Overview

Metric	Value
Sessions	2,329
Total turns	129,193
Tool calls	119,146
API calls	109,371
Total tokens	33.7B (incl. cache)
Avg context per call	81,881 tokens
Cache savings	85.3%

Tool Distribution

Tool	Calls	Share
Bash	48,050	40.3%
Edit	23,369	19.6%
Read	21,818	18.3%
Glob	8,212	6.9%
Grep	6,076	5.1%
TodoWrite	4,770	4.0%
Write	3,791	3.2%
Task	895	0.8%
Other	2,165	1.8%

Bash dominates at 40%. This is partially inherent (Rust repos need cargo for build/test/clippy/fmt), but the sequential chaining patterns suggest significant waste.

Per-Repo Breakdown

Cost and Volume

Repo	Sessions	Cost	$/session	Turns	Avg ctx
unshape	295	$683	$2.31	19,931	83,255
normalize	162	$678	$4.18	18,502	89,228
hologram	434	$369	$0.85	12,635	76,997
io	321	$332	$1.03	12,123	71,438
sketchpad	41	$319	$7.78	7,769	96,781
ooxml	58	$291	$5.01	7,933	91,974
wick	117	$225	$1.92	6,278	84,418
parents	237	$201	$0.85	8,920	62,981
server-less	59	$181	$3.07	5,553	84,800
zone	124	$150	$1.21	4,525	78,821
lotus	62	$148	$2.39	4,036	92,050
rescribe	30	$137	$4.58	3,790	89,516
paraphase	32	$135	$4.22	4,381	81,849
moonlet	79	$78	$0.99	2,556	73,457
aspect	60	$75	$1.25	3,565	65,724
pad	24	$43	$1.81	1,741	75,665
hub	32	$36	$1.13	1,089	87,971
portals	47	$28	$0.59	1,146	64,596
myenv	13	$20	$1.55	541	76,797
playmate	22	$10	$0.45	341	61,565
keybinds	15	$9	$0.60	578	47,928
herbarium	11	$7	$0.60	213	68,174
concord	9	$11	$1.26	362	80,709
pteraworld	6	$5	$0.80	329	48,936

Key observations:

• sketchpad ($7.78/session) and ooxml ($5.01/session) are the most expensive per-session. Both have high avg context (97K, 92K) suggesting large codebases that inflate every API call.
• normalize and rescribe are also expensive per-session ($4.18, $4.58) with high context (89K each).
• hologram and parents are cheap per-session but have the most sessions (434, 237), so their total spend adds up.

Pain Point 1: Sequential Bash Chains

The dominant tool pattern in nearly every repo is long Bash -> Bash -> Bash -> Bash -> Bash chains. These represent the agent running commands one-at-a-time when many could be parallelized.

Repo	5+ Bash chains	% of all tool patterns
unshape	4,548	dominant
parents	4,228	dominant
server-less	2,234	dominant
paraphase	1,873	dominant
io	1,583	dominant
normalize	1,358	dominant
zone	1,013	dominant
wick	1,002	dominant
moonlet	959	dominant
ooxml	1,105	dominant

What's happening: The agent runs cargo build, waits, runs cargo test, waits, runs cargo clippy, waits, runs cargo fmt --check, waits. Each is a separate API round-trip. In the normalize session analyzed, this pattern accounted for 217 parallelizable API calls in a single session.

Fix: CLAUDE.md guidance to batch cargo commands: cargo fmt -- --check && cargo clippy -- -D warnings && cargo test in a single Bash call.

Pain Point 2: Glob Spam

Several repos show chains of 5+ consecutive Glob calls as a top pattern:

Repo	5+ Glob chains
normalize	1,012
pad	982
paraphase	969
hologram	893
wick	767
io	699
rescribe	681

What's happening: The agent is doing file discovery one glob pattern at a time. For example, searching for *.rs, then *.toml, then *.md, then *.ts, etc., each as a separate tool call.

Fix: Use the Task/Explore agent for broad file discovery. When doing targeted searches, batch multiple globs into a single call or use normalize view <dir> for structural overview.

Pain Point 3: Repeated File Thrashing

Individual session analysis reveals excessive re-reading and re-editing of the same files. The worst example from a wick session:

File	Reads	Edits	Total touches
dew-linalg/Cargo.toml	6	14	20
dew-cond/Cargo.toml	6	11	17
dew-quaternion/Cargo.toml	5	12	17
dew-complex/Cargo.toml	5	12	17
dew-scalar/Cargo.toml	5	11	16
dew-linalg/src/lib.rs	6	8	14
dew-scalar/src/lib.rs	7	7	14

What's happening: Incremental edits that break the build, leading to read-edit-build-fail-read-fix cycles. The agent makes a change to one Cargo.toml, finds the build broken, fixes it, but then needs to make the same change to 5 other Cargo.toml files -- one at a time.

Fix: CLAUDE.md guidance to plan multi-file changes upfront. When making the same change across crate Cargo.toml files, do all edits before building.

Pain Point 4: Context Pressure

Sessions regularly approach the ~155K token context limit:

Session	Repo	Peak context	Effect
28d418c3	wick	141.7K	Approaching limit
6b847726	normalize	131.0K	Approaching limit
(aggregate)	sketchpad	avg 96.8K	Consistently high
(aggregate)	ooxml	avg 92.0K	Consistently high
(aggregate)	lotus	avg 92.1K	Consistently high

What's happening: Long sessions accumulate context from tool results (especially cargo build output, large file reads). The normalize session showed context growth from 10K to 131K over 283 turns.

Fix: Use normalize view to get structural outlines instead of reading entire files. Use offset/limit for targeted reads. In long sessions, prefer the Task agent to offload exploration.

Pain Point 5: Error-Driven Rework

"Corrections & Apologies" (the agent saying "Let me fix") correlate with wasted cycles:

Session	Repo	Corrections	Build errors
28d418c3	wick	6	8 command failures
6b847726	normalize	4	17 command failures
b70ecf3d	unshape	1	file-not-found, token limit

Common error patterns across all sessions:

• cargo fmt failures after edits - agent doesn't run fmt proactively
• Compilation errors from incremental changes - untested assumptions
• "File does not exist" - agent guessing at file paths
• "File has not been read yet" - agent trying to write without reading first
• Token limit exceeded - reading large files without offset/limit

Pain Point 6: Underused Tools

Tool	Calls	Share	Note
Task (subagents)	895	0.8%	Drastically underused for exploration
AskUserQuestion	139	0.1%	Agent rarely asks for clarification
EnterPlanMode	158	0.1%	Rarely plans before acting
Grep	6,076	5.1%	Often Bash grep/rg used instead

The agent defaults to sequential Bash commands for exploration when Task/Explore agents would be more efficient and context-friendly.

Actionable Changes

Based on this analysis, the following changes have been rolled out across ecosystem repos:

1. Auto-Format Pre-Commit Hook

Changed: Pre-commit hooks across all Rust repos now run cargo fmt --all (apply) instead of cargo fmt --all -- --check (check-only). The hook also stages the formatted changes with git add -u.

Why: A huge portion of the Bash chains were cargo fmt --check → fail → cargo fmt → re-stage → retry. The agent no longer needs to think about formatting at all — it's handled at commit time.

Affected repos: All Rust repos in the ecosystem (normalize, unshape, wick, moonlet, paraphase, rescribe, concord, reincarnate, server-less, playmate, portals, zone, myenv, interconnect, ooxml). Scaffolding template updated.

2. Cargo Command Batching

Added ## Workflow section to all Rust repo CLAUDE.md files:

Batch cargo commands to minimize round-trips:
  cargo clippy --all-targets --all-features -- -D warnings && cargo test
After editing multiple files, run the full check once — not after each edit.

3. Multi-File Edit Discipline

Added to all multi-crate workspace CLAUDE.md files:

When making the same change across multiple crates, edit all files first, then build once.

4. Structural Exploration

Added normalize view availability note to all repo CLAUDE.md files. Not prescriptive — just making it known as an option for structural outlines of files and directories.

5. Scaffolding Template

Updated scaffolding/CLAUDE.md and scaffolding/.githooks/pre-commit to include all the above for new repos.

---

Analysis performed with normalize sessions across ~/.claude/projects/. Data covers all sessions through January 2026.