Read your Claude Code and Codex session logs locally. See where the tokens went — by bucket, by request, with prompt-cache and API-equivalent cost.

English · 简体中文

You're running Claude Code or Codex. The context bar creeps toward 80%, the agent slows down, and you have no idea what's actually in there. Tool results? An MCP server you forgot you enabled? Six months of summaries?

cw-inspect reads the same local JSONL transcripts the CLI already writes, then tells you:

• Exact input / cached / cache-write / cache-read / output / reasoning tokens (from the recorded usage field — not estimated).
• Approximate attribution by bucket: tool results, assistant messages, tool calls, system prompts, summaries, hooks, skills, MCP tools.
• API-equivalent cost per request and per session, using public model prices.

No API key. No network calls. Just the files already on your disk.

⭐ If this saves you a "where did my context go?" investigation, a star helps other Claude Code and Codex users find it.

A one-line, colored bar you can drop into Claude Code's statusLine or print from your shell. The widest segment is usually tool — that's the signal.

Screenshots are representative terminal renderings. Live values and table widths vary by session, model, terminal width, and current transcript contents.

Markdown by default, JSON with --json. Pipe it anywhere.

Install the CLI from the latest release:

python3 -m pip install \
  "context-window-inspector @ https://github.com/androidZzT/context-window-inspector/releases/download/v0.1.0/context_window_inspector-0.1.0-py3-none-any.whl"

Or install the CLI from a checkout:

git clone https://github.com/androidZzT/context-window-inspector.git
cd context-window-inspector
python3 -m pip install -e .

Requires Python 3.10+. Zero runtime dependencies.

The wheel installs the cw-inspect CLI. The Codex plugin below is a repo-local marketplace plugin, so keep a source checkout when you want Codex to load it.

Inspect the most recent session:

cw-inspect claude --latest
cw-inspect codex  --latest

Pick a specific session by id prefix or path:

cw-inspect claude --session 137d4ea4
cw-inspect codex  --session ~/.codex/sessions/.../rollout-XXXX.jsonl

Useful flags:

Print the compact bar yourself:

cw-inspect statusline claude --latest
cw-inspect statusline codex  --latest --no-color --ascii

Wire it into the CLI you already use:

cw-inspect install-statusline claude   # wraps ~/.claude/statusline.sh
cw-inspect install-statusline codex    # writes ~/.codex/statusline-cwi.sh + enables native items
cw-inspect install-statusline all

Notes:

• Claude Code supports command-driven status lines, so the installer wraps your existing statusline.sh and appends the context split.
• Codex currently only accepts built-in [tui].status_line identifiers. The installer enables native context-used and used-tokens, and writes ~/.codex/statusline-cwi.sh for the detailed breakdown — call it from your terminal prompt or a tmux status bar. Restart Codex (or open /statusline) after installing.

plugins/context-window-inspector/ ships a Codex plugin that exposes a get_codex_context_window MCP tool, so Codex itself can answer "what's filling my context window?" mid-session.

The plugin is loaded from this repository checkout; it is not installed into Codex by the Python wheel.

Add to ~/.codex/config.toml:

[marketplaces.context-window-inspector]
source_type = "local"
source = "/path/to/context-window-inspector"

[plugins."context-window-inspector@context-window-inspector"]
enabled = true

Restart Codex. Ask the agent. It can return Markdown, the compact statusline string, or JSON.

Two layers, kept separate on purpose:

1. Exact tokens — from the provider, not us.

If the provider didn't record it, we don't show it.

2. Bucket attribution — observable payload, not a tokenizer guess.

For each transcript record we count bytes and characters and bin them into buckets (tool_results, assistant_messages, tool_calls, system_or_base_instructions, summaries, developer_or_mode_instructions, mcp_or_dynamic_tools, hooks, skills, user_messages). The split percentages allocate the latest exact token count across buckets weighted by byte size, so they're approximate by design. We don't run a tokenizer.

For Claude Code statuslines, the context percentage uses Claude Code's own input-only formula: input_tokens + cache_creation_input_tokens + cache_read_input_tokens.

Hook attribution is deliberately narrow. Hook execution logs, echoed tool inputs, and debug stdout are not counted as context. Only additionalContext, systemMessage, stdout from context-bearing prompt/session events, and blocking errors land in the hooks bucket.

Reports include an API-equivalent cost when a public price is configured for the model. Read this as "what would this look like on the API meter," not as "what your subscription will be billed."

• OpenAI / Codex: cached_input_tokens is priced as the cached subset of input_tokens. reasoning_output_tokens is treated as a detail of output_tokens, not an extra charge.
• Claude: cache_creation_input_tokens, cache_read_input_tokens, and output_tokens are priced separately. When Claude reports cache_creation.ephemeral_5m_input_tokens and ephemeral_1h_input_tokens, each duration uses its own cache-write rate.

Public price tables live in src/context_window_inspector/pricing.py. PRs welcome when prices change.

• Not a billing system. Estimates can drift from your invoice.
• Not a tokenizer. Bucket splits are byte-weighted approximations.
• Not a cloud service. Everything runs against files on your machine.

src/context_window_inspector/
  cli.py          # argparse entry point
  claude.py       # Claude Code JSONL parser
  codex.py        # Codex JSONL parser
  models.py       # ExactUsage / SessionReport / bucket types
  pricing.py      # public price tables, cost estimator
  reporting.py    # markdown + JSON renderers
  statusline.py   # compact bar renderer
  install.py      # statusline installers
plugins/
  context-window-inspector/   # Codex MCP plugin
tests/                          # parser, pricing, statusline, plugin tests

python3 -m pytest
PYTHONPATH=src python3 -m context_window_inspector codex --latest

See AGENTS.md for module conventions. Keep parsing, aggregation, and rendering in separate modules; provider-specific transcript handling stays out of the CLI layer.

Issues and PRs welcome. If a parser drops a field your provider records, open an issue with a redacted snippet and which CLI version produced it.

MIT — see LICENSE.