ctxlens

A flamegraph for your agent's context window. ctxlens parses AI agent session transcripts and shows you exactly what is eating your context per turn, flags the waste, and tells you what to cut.

Why

Agents get slow, expensive, and dumb when their context window fills with junk: the same file read six times, a 12k-token tool result that mattered for one turn, tool schemas re-sent on every step. Token dashboards tell you the bill. ctxlens tells you where the bytes went and what to delete.

It works offline with a deterministic heuristic tokenizer (no network, no heavy deps), and upgrades to exact counts automatically when tiktoken is installed.

Quickstart

pip install ctxlens-cli          # core
pip install "ctxlens[tiktoken]"   # optional exact token counts

ctxlens analyze session.jsonl
ctxlens report session.jsonl --html -o report.html
ctxlens diff before.jsonl after.jsonl

Point it at a real Claude Code session, or pipe a transcript in:

ctxlens analyze ~/.claude/projects/<slug>/<session>.jsonl
cat session.json | ctxlens analyze - --json

Example output

╭─ ctxlens ────────────────────────────────────╮
│ Source   session.jsonl                        │
│ Format   claude-code-jsonl   Tokenizer heuristic │
│ Tokens   12,481   Turns 14   High-water 12,481 │
│ Waste    4,932 tokens (39.5%)                 │
╰───────────────────────────────────────────────╯
Context composition by segment
 Segment       Tokens     %  Msgs  Share
 tool result    6,204  49.7    22  ██████████████·······
 assistant      2,110  16.9    14  ██████···············
 system         1,540  12.3     1  ████·················
 ...
Context growth
 cumulative  ▁▁▂▂▃▄▄▅▆▆▇▇██  peak 12,481
 per-turn    ▂█▃▂▅▂▁▂▇▂▁▃▂▁  max 3,204
Recommendations
 [HIGH] Repeated content wastes tokens  (~2,410 tok)
     'Read:file_path=config.py' appears 6 times (turns [2, 5, 7, 9, 11, 13]) ...

Supported formats

ctxlens auto-detects the format; override with --format.

Every message is attributed to a segment: system, tool_definitions, user, assistant, thinking, tool_call, tool_result.

Recommendations explained

The engine is rule-based and specific, not generic advice:

• Tool results dominate — flagged when tool outputs are >=30% of context.
• Repeated content — the same file/tool result (by reference or by exact body) appearing more than once; every copy after the first is wasted.
• Stale tool outputs — an older result superseded by a newer one for the same target still occupies context.
• Oversized tool definitions — tool schemas above budget, paid every turn.
• System prompt weight and single biggest consumer callouts.

Each recommendation carries a severity and an estimated token saving.

Waste report

waste_ratio = total_waste / total_tokens, where total waste sums duplicate tokens, tool-result bloat (tokens above --tool-result-cap), stale tool outputs, and tool-definition overage (above --tool-def-budget).

CI usage

Fail a build when a captured agent session wastes too much context:

ctxlens analyze session.jsonl --fail-over-ratio 0.30

Exit codes: 0 OK, 2 threshold exceeded, 1 error. Emit machine-readable output with --json, or diff a baseline against a candidate in CI with ctxlens diff baseline.jsonl candidate.jsonl --json.

Tokenizers

• heuristic (default fallback): deterministic, dependency-free, great for relative profiling and CI.
• tiktoken: exact BPE counts when installed. --tokenizer auto prefers it.

Contributing

git clone https://github.com/AgentPostmortem/Ctxlens
cd ctxlens
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
ruff check .
pytest

Issues and PRs welcome. Adding a parser? Implement Parser.sniff and Parser.parse, register it in ctxlens/parsers/__init__.py, and drop a fixture in tests/fixtures/.

License

MIT © royalpinto007