Context Mode cut a 331 KB log to 571 bytes. Here's the catch.
We measured Context Mode on four fixtures: a 331 KB noisy log dropped to 571 bytes without losing the three failing test names.
By David Discua
Measured in the lab
- 99.8%
- bytes saved
- 99.7%
- tokens saved
Bars on a log scale.
- test_auth_flow found
- test_payment_webhook found
- test_inventory_sync found
- 11 MCP tools discovered
From reports/bench.json · run on October 13, 2026. cl100k_base tokens; context-mode 1.0.169; 4 fixtures; harness discovered 11 tools
Where it sits
01/03Your tool dumps a three hundred thirty-one kilobyte log.
We ran Context Mode's MCP sandbox over four fixtures with no API keys. A synthetic 331 KB CI log came back as 571 bytes, 47,986 tokens down to 152, and the three failing test names were still in the output. The catch is that the squeeze depends on the fixture, so read the percentages as our lab numbers from 2026-10-13 and re-run the bench on your own logs.
The catch
The squeeze depends on the fixture. Our tokenizer is cl100k_base, our pin is context-mode@1.0.169, and our noisy log is synthetic. The percentages are a lab measurement of those inputs. Re-run the bench on your own logs before you redesign an agent around them.
What we ran (keyless)
bun install && bun run bench
Headline numbers from reports/bench.json (run on October 13, 2026):
| Fixture | Raw bytes | CM bytes | Byte Δ | Raw tokens | CM tokens |
|---|---|---|---|---|---|
| noisy_log | 331,109 | 571 | 99.8% | 47,986 | 152 |
| json_api | 2,180,338 | 923 | 100.0% | 837,849 | 308 |
| git_stat | 374,139 | 643 | 99.8% | 137,773 | 197 |
| dir_listing | 235,015 | 433 | 99.8% | 97,850 | 127 |
On the noisy log, all three expected failure names survived the squeeze (test_auth_flow, test_payment_webhook, test_inventory_sync). The harness discovered 11 MCP tools.
How it is wired
Context Mode sits between tool output and the model. The agent still gets the failure names and a searchable index, and the raw log stays out of the prompt.
Minilab (open source)
Upstream: mksglu/context-mode
The bench writes reports/bench.json, which is where every number above comes from.
Run it yourself
SkilledDragonCorp/minilab-context-mode
Clone, install and run the report with no API keys. The numbers above are what that run writes.
bun install && bun run benchSkilledDragon is not affiliated with context-mode or the projects named. Numbers are from our own lab run; re-run the bench before relying on them. Marks belong to their owners.
All labs