e2e-results: 2026-07-30 testbench run (claude harness)

[jstirnaman]

Summary

Adds the 2026-07-30 agent e2e testbench run for the runbook in #69's AGENT_E2E_TESTS.md, executed via the docs-agent testbench harness (Claude only, per-case docker-compose live instances).

Tested against InfluxDB 3 Core 3.10.3 and InfluxDB 3 Enterprise 3.11.0-rc.1.

Findings

• 11/12 runs passed grading against each case's expected behavior (correct tool choice, bounded queries, correct read/write behavior per server mode).
• 1 needs_review: ro-absent-table — the agent's answer content was correct (found: false, no query_sql run), but the transcript shows it called Bash ({"command":"true","description":"noop"}) and ToolSearch, violating the case's explicit "Do not use shell commands" constraint. Left as needs_review rather than a hard fail — the results schema has no failed enum value (only passed/interrupted/needs_review).
• Tool defect surfaced during grading: eff-broad-schema-cpu (default/operator mode) sent a query_sql call with the query text under query instead of the required q param, got MCP error -32603 (invalid_type, path: ["q"]), then retried correctly with q. Filed as query_sql: agent naturally sends 'query' param, tool requires 'q' — causes avoidable failed call #91 — the natural param name for a SQL-query tool is query, not q, so this is likely to recur across agents/harnesses even though this run recovered.
• Efficiency comparison (ro-wildcard-field-low vs -high, reasoning low vs high): identical 4-call tool path and correctness, but high-effort reasoning cost 2959 vs 1895 tokens for no behavioral difference.
• Mode comparison (eff-broad-schema-cpu, ro vs default): default mode used get_measurements instead of list_tables, and was both less efficient (12 vs 11 tool calls) and less reliable (1 failed call vs 0) for the same task.

References

• query_sql: agent naturally sends 'query' param, tool requires 'q' — causes avoidable failed call #91 (malformed query_sql call, missing q)
• Feat: Implement Phase 1 read-only MCP plan #69 (AGENT_E2E_TESTS.md runbook this run executes)
• docs-tooling docs-agent/testbench/README.md (harness design, grading process)

Test plan

• AGENT_E2E_RESULTS schema validation passed (node docs-agent/testbench/schemas/validate-e2e-results.mjs in docs-tooling, 12/12 runs valid) before this file was copied over