[dan-petty]

A call to a hosted gateway group (claude-sonnet-5-5, claude-haiku-5-5, deepseek-v4.1-flash-1017) was recorded as a local call at $0, and its tokens counted as local savings. The hosted groups set no api_base, so the gateway reply names a deployment model and no backend, and is_local judged the call by the gateway host, a local host. A reply of the gateway provider that names a deployment model and no backend is now hosted, and the lifetime report groups the served usage by the deployment model. Claude and Gemini calls are now priced by genai-prices. genai-prices holds no price for deepseek-v4.1-flash-1017, so the warning for an unpriced hosted model now names the command devops ai cost set-price. Rows written before this change keep the source local.

Closes #3656

Change: openspec/changes/archive/2026-11-05-3656-hosted-gateway-priced-local/

[dan-petty]

Fail-to-pass: devops ci fail-to-pass --base origin/main reports 8 failed on base (8 failed and 143 passed in tests/ai/benchmark/test_k8s_llm_gateway.py, tests/ai/spend/test_ai_spend_ledger.py, tests/ai/spend/test_ai_spend_pricing.py; 0 errored). The CI check Fail-to-Pass Check passes.

The 8 tests: is_local of a reply with a deployment model and no backend, the price of the Claude and Gemini groups behind a loopback gateway (3 cases), the warning and the price override of the DeepSeek group, the lifetime report, and the two older tests that pin the text of the warning. test_every_local_deployment_names_its_backend_so_the_ledger_counts_it_as_local passes on base on purpose: the gateway config is already correct, and the test fails if a local deployment loses its api_base.

Full local gate (devops ci --check): lint, format, typecheck, security, docs, openspec, build and the other 19 gates pass. The test gate has 4 failures and 1 error, in tests/crypto/test_ssh.py, tests/telemetry/test_config_env_overrides.py, tests/commands/test_devcontainer_install_tools_dry_run.py and tests/ai/test_tool_executor_contract.py. They fail the same way with this change stashed, so they come from the state of the dev container (an installed popeye, the name of the checkout directory, a /tmp/config.yaml, tool install status), not from this change.

What this does not do: it sets no price for deepseek-v4.1-flash-1017. The three known prices differ (gateway config 0.05 and 0.192 USD per million tokens, OpenRouter list 0.30 and 1.20, genai-prices deepseek/deepseek-v4.1-flash 0.0352 to 0.0448 and 0.08). A person sets the price with devops ai cost set-price openrouter/deepseek/deepseek-v4.1-flash-1017:floor <prompt_rate> <completion_rate>. Until then the group has the source unknown and one warning.

Ledger rows that were written before this change keep the source local. The caching follow-up is #3658.

[dan-petty]

Scope note. This pull request fixes the calls whose gateway reply names a deployment model: the chat_dispatch rows. The agent_run rows of the same ledger (301 rows, 20.18 M prompt tokens against 10.53 M for chat_dispatch) record no reply header, so this change does not reach them. They are #3665. Merging this pull request does not close that problem.