[zwanski-start]

Closes #107

Problem

With Ollama (or any custom ANTHROPIC_BASE_URL) configured via .env, the agent still requested claude-opus-4-5-20251101 and got a 404. The model passed to the CLI comes from the stored UI setting (api_model or model, falling back to DEFAULT_MODEL). That value is a full Claude ID. The Claude CLI applies ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU}_MODEL only to the tier aliases (opus/sonnet/haiku), not to explicit IDs, so the Claude ID went straight to Ollama. Switching to Sonnet in the UI hit the same problem.

Fix

• client.py: new resolve_model_for_endpoint(model, sdk_env). When ANTHROPIC_BASE_URL is set (and Vertex is off) and the model is a claude-* ID, it is replaced with the matching tier override (opus → ANTHROPIC_DEFAULT_OPUS_MODEL, sonnet → ..._SONNET_..., haiku → ..._HAIKU_...). Non-Claude model names, missing or blank overrides, Vertex, and the plain Anthropic API are left alone. It's applied in create_client, the single point where the model reaches the SDK, so the single-agent, parallel, testing and scheduled paths are all covered.
• The Ollama/Alternative API startup log lines now print the model actually used, which makes this kind of mismatch easy to spot.

Verification

• Added TestResolveModelForEndpoint (9 cases) to test_client.py: venv/bin/python -m pytest test_client.py passes (30 tests).
• ruff check . is clean and python test_security.py passes.
• mypy client.py reports only a pre-existing unused-ignore on the effort= line (it depends on the local SDK version and is unrelated).

Assumptions

• Settings-UI providers (Ollama/GLM/etc. selected in the UI) already pass api_model and set all three overrides, so this only changes behavior for .env-configured endpoints, or when a Claude ID still reaches a custom endpoint.
• Proxies that really serve Claude models usually set no tier overrides, or set them to Claude IDs, so they keep working.

🤖 Generated with Claude Code

[zwanski-start]

🔍 QA report: ✅ Passed

Tester: Ken (cubefarm agent, in a fresh session) · Round: 1 · Author: Dennis

The fix works: with Ollama set up through .env (ANTHROPIC_BASE_URL plus tier overrides), a stored Claude model ID such as claude-opus-4-5-20251101, or a Sonnet pick, is now replaced with the matching ANTHROPIC_DEFAULT_*_MODEL. Before, it was sent as is and got a 404. A side-by-side check of create_client confirms master sends the Claude ID and this branch sends qwen2.5-coder. Plain Anthropic, Vertex and Settings-UI providers are unchanged. Lint, tests, the UI build and a server smoke test show no regressions; the remaining test and mypy failures are the same on master.

	Check	Details
✅	Issue #107 repro: .env Ollama setup with a stored Claude opus ID	Called create_client with ClaudeSDKClient mocked and get_effective_sdk_env patched to the issue's env. Master passed model=claude-opus-4-5-20251101; the PR passes qwen2.5-coder and logs "Ollama Mode: Using local model 'qwen2.5-coder'".
✅	Issue repro: picking Sonnet in the UI	claude-sonnet-4-6 now resolves to the ANTHROPIC_DEFAULT_SONNET_MODEL value. On master it was sent unchanged.
✅	No regression for plain Anthropic, Vertex and Settings-UI Ollama	No base URL: claude-opus-4-7 unchanged. CLAUDE_CODE_USE_VERTEX=1: unchanged. Settings-UI Ollama with api_model qwen3-coder: unchanged.
✅	Edge case: missing tier override	claude-haiku-4-5 with no haiku override passes through unchanged, as the PR states. The new log line shows that model name, so a mismatch is visible.
✅	Code review	Small pure function, applied in create_client, which every agent path uses. Log lines print only the model name, no secrets. Security allowlist and middleware untouched. 9 new unit tests cover the cases.
✅	Python test suite compared with master	153 passed, 21 failed (test_devserver_security.py). The failure list is identical with master's client.py and test_client.py, so the failures come from elsewhere.
✅	Server boots, REST routes and WebSocket	uvicorn on 127.0.0.1:5345 with an isolated HOME. GET / 200, /api/settings 200, /api/projects 200, POST /api/projects with {} returns 422. Creating a project and a feature returned 200. The dashboard showed the WebSocket as Live.
✅	UI renders in the browser	Dashboard and dependency graph rendered with the test feature. No console errors or warnings. This is a backend-only change, so the UI e2e suite was not needed.

🧪 Commands run

Command	Result
ruff check .	All checks passed
mypy .	24 errors in 9 files, same as master (the client.py unused-ignore is pre-existing)
python -m pytest -q test_client.py	30 passed
python -m pytest -q	153 passed, 21 failed (same 21 failures on master)
python test_security.py	ALL TESTS PASSED
python test_security_integration.py	ALL INTEGRATION TESTS PASSED
python repro.py (create_client with SDK mocked, PR vs master client.py)	PR: Claude IDs mapped to qwen2.5-coder; master: Claude IDs sent unchanged (bug reproduced)
npm --prefix ui run build	success
npm --prefix ui run lint	0 errors, 6 pre-existing warnings
uvicorn server.main:app --port 5345 + curl route probes	200/200/200/422 as expected

No browser screenshots were taken in this round.

_{Posted by cubefarm · merges automatically once GitHub's checks pass}