Context efficiency

Tokenizer: exact o200k_base encoding · 3 measured repetitions after warmup.

scenario	surface	tool-definition tokens	total task-context tokens	MCP body KB (request + response)	p50 ms	p95 ms
small (8 tools, 1 KB result)	flat_mcp	1313	1746	8.4	17.41	18.16
small (8 tools, 1 KB result)	compact_mcp	1026	2465	11.8	19.89	21.02
small (8 tools, 1 KB result)	passport_cli	0	1165	17.8	336.69	342.92
medium (48 tools, 8 KB result)	flat_mcp	4933	7356	32.3	20.71	23.32
medium (48 tools, 8 KB result)	compact_mcp	1106	4825	21.6	19.68	20.1
medium (48 tools, 8 KB result)	passport_cli	0	3389	28.1	350.99	356.59
large (160 tools, 32 KB result)	flat_mcp	15069	24203	104.2	37.69	40.26
large (160 tools, 32 KB result)	compact_mcp	1212	11680	49.7	27.17	27.55
large (160 tools, 32 KB result)	passport_cli	0	10086	56.8	361.85	365.78

Context ratios versus flat MCP

scenario	compact definitions	compact total	CLI total
small	78.1%	141.2%	66.7%
medium	22.4%	65.6%	46.1%
large	8%	48.3%	41.7%

Tool-definition and task-context tokens are exact for the o200k_base text encoding, but host-specific MCP wrappers and any generic shell-tool schema are excluded equally. They are not a billing estimate for every model.
MCP wire figures are JSON HTTP body bytes only. Headers are excluded because bearer length is noise and credentials must never enter output. CLI auth exchange bytes are recorded separately in JSON.
CLI context assumes the agent already knows the stable Passport command contract. One-time installation/help text is excluded; every per-task command and stdout byte is included.
Latency includes localhost gateway/upstream work. CLI includes two Node subprocesses, two credential exchanges, and two MCP sessions. Latency is observational and deliberately has no pass/fail threshold.
Compact/CLI search returns at most ten matching definitions. This task searches across the permitted fleet, selects the namespaced result, and performs one governed call.

Canaries: 33/33 passed.