Feat/claude model selection by mistercap179 · Pull Request #11 · emanuilo/memclaw

mistercap179 · 2026-09-21T15:34:26Z

Summary

Adds Claude model and effort-level selection to Memclaw.

Until now, the Claude backend always ran a hardcoded claude-sonnet-4-6 with no way to change the model or how deeply it thinks. The setup wizard now asks both questions, and the answers are stored as ordinary config values.

What changed

Live model discovery

memclaw/anthropic_models.py

Fetches the available models from Anthropic's /v1/models using the credential the user just entered, so a newly released model appears without a Memclaw release. Both credential types are supported (OAuth bearer token and x-api-key); they are not interchangeable, so the auth header is chosen per mode.
Reads each model's effort capability tree and keeps only the levels the installed claude-agent-sdk accepts (low / medium / high / xhigh / max), so a level a future API adds drops out on its own instead of reaching the SDK.
Models are sorted newest first, and the list is cached in ~/.memclaw/models_cache.json for 24h so repeated memclaw configure runs don't re-hit the API. A missing, stale, or malformed cache silently falls back to the network.

Wizard

memclaw/backends/claude.py

After the credential step, a numbered model picker is shown with the current model preselected, followed by an effort picker — but only for models that report effort support.
For a model with no effort support, the effort question is skipped and any level left over from a previous choice is dropped, so it cannot be sent to a model that would reject it.
Every failure path (no network, bad key, empty list) warns once and keeps the current model. Picking a model is a convenience and must never be the reason memclaw configure cannot finish.

Config and runtime

Adds ANTHROPIC_MODEL and ANTHROPIC_EFFORT settings, resolved from the config file or the environment like every other value.
Both settings are optional:
- An unset model falls back to the built-in default.
- An unset or unrecognised effort level makes the SDK omit the flag entirely.
The backend passes the resolved model and effort to ClaudeAgentOptions.
memclaw status shows the active model and effort, but only when Claude is the active backend — printing them under Cursor would name a model that isn't going to run.

Dependency

Bumps claude-agent-sdk to >=0.2.139,<0.3, the first version exposing ClaudeAgentOptions.effort.

Tests

New coverage for:

Model fetching and parsing
Effort filtering
Cache read/write/expiry
Model and effort resolution
Wizard picker, including skip-effort and fetch-failure paths
Backend tests isolated from the developer's own environment so a local ANTHROPIC_MODEL cannot change the result

Notes

Defaults are unchanged: an existing install that accepts every prompt keeps running the same model it ran before.

README documents both variables and the new wizard steps.

The previous range capped at 0.1.x, which has no `effort` option. 0.2.x exposes it as Optional[Literal["low","medium","high","xhigh","max"]], which the upcoming model/effort selection needs. Verified against the installed 0.2.152 that every ClaudeAgentOptions field this project uses still exists: env, model, system_prompt, setting_sources, mcp_servers, allowed_tools, disallowed_tools, permission_mode, max_turns. Test suite is unchanged by the bump: 178 passed, 4 failed. The same four failures (tests/backends/test_mcp_bridge.py) occur on main before this commit. They are Windows-only and unrelated: mcp_bridge.py passes fd= to uvicorn, which resolves it via socket.fromfd(..., AF_UNIX, ...), and AF_UNIX does not exist on Windows.

Adds memclaw/anthropic_models.py, which asks /v1/models what the configured credential can reach and reads each model's effort support off its capability tree. The setup wizard will build its model picker from this, so a model Anthropic releases needs no code change here. Shaped after openai_health.py: probe an API, return a small dataclass report, leave presentation to the caller. Notes on the pieces that are easy to get wrong: - Auth headers differ by credential. An OAuth token sent as x-api-key is rejected 401, so subscription and API-key modes build separate headers. - Effort levels are filtered through SDK_EFFORT_LEVELS, mirroring what ClaudeAgentOptions.effort accepts. A level Anthropic ships before our SDK understands it is dropped rather than passed through. - Results are cached for 24h in ~/.memclaw/models_cache.json. A missing, malformed or future-dated cache falls back to the network instead of raising; the wizard must never break over a cache file. - The endpoint pages at 20 by default, so limit=100 is requested. Verified against the live API with a subscription token: 11 models, correctly ordered newest first. Effort support turns out to be graded rather than all-or-nothing, which the capability tree handles for free: Sonnet 4.6 supports low/medium/high/max but not xhigh, Opus 4.5 stops at high, and Haiku 4.5 and Sonnet 4.5 report no effort support at all. 23 new tests, all HTTP mocked; the suite still needs no keys and no network. Full run: 201 passed, plus the 4 pre-existing Windows-only mcp_bridge failures that also fail on main.

Two new MemclawConfig fields alongside cursor_model, populated from the environment the same way CURSOR_MODEL already is. Empty means "use the built-in default", so nothing changes for an existing install until the wizard starts writing these keys. They sit in their own Claude block rather than inside the Cursor one, to keep the two backends visually separate. 204 passed, plus the 4 pre-existing Windows-only mcp_bridge failures.

The Claude backend's wizard_setup now follows the credential question with a model picker built from the live /v1/models list, and an effort picker holding only the levels that model actually accepts. A model that supports no effort skips the second question entirely and drops any level left over from an earlier choice, so an unsupported value can never reach the SDK. Fetching the list is a convenience, not a requirement: any failure prints one warning and keeps the current model, so the wizard always finishes.

Both run_one_shot and run_turn now take the model and effort off the config instead of the hardcoded constant. _MODEL stays as the fallback for an unset value, and an effort the SDK does not recognise resolves to None, which makes the SDK omit the flag rather than pass it to the CLI. Both are resolved once in __init__, mirroring how the Cursor backend resolves its model. Verified against the real SDK, not just the mocks: a one-shot turn succeeds with effort=high on Sonnet 4.6 and with effort unset on Haiku 4.5.

The panel now reports which model a turn will actually use and which effort level goes with it, resolved through the same helpers the backend uses, so status cannot drift from runtime. Both lines are skipped when Cursor is the active backend — naming a Claude model there would describe something that is not going to run. "default" means no level is set and the model decides.

Adds the cases the spec asks for that the backend tests did not yet have: an empty ANTHROPIC_MODEL falls back to the built-in default, an effort level the SDK does not know is ignored, both values reach ClaudeAgentOptions, and a failed model fetch warns once instead of raising out of the wizard.

The autouse fixture already stripped the credential env vars, but not ANTHROPIC_MODEL/ANTHROPIC_EFFORT. Importing memclaw.config loads ~/.memclaw/.env into os.environ, so once a developer had run the wizard, the three default-fallback tests asserted against that developer's own model instead of the built-in default.

anthropic_models.py and backends/claude.py imported from each other, so both sides needed in-function imports. fetch_models() now takes the auth mode and credential as plain strings, like probe_openai(api_key), so the module no longer needs anything from claude.py and lives next to it as backends/claude_models.py. All imports in claude.py are back at the top. This also removes _probe_config, which built a whole MemclawConfig (and its folders and files) just to carry one credential. The model-list cache is gone: it could serve a list fetched with a previous credential, and the picker only runs from `memclaw configure`. The OAuth beta header is dropped too, since /v1/models doesn't need it.

The Claude CLI reads ANTHROPIC_MODEL itself, so someone who exports it for Claude Code would silently change Memclaw's model too. The new names match CURSOR_MODEL, and a test pins that ANTHROPIC_MODEL is ignored.

The status command checked for the Claude backend by name and imported its private helpers, which base.py says nothing outside a backend should need. Backends now implement status_rows(config) -> [(label, value)]: Claude returns model and effort, Cursor returns its model, and status prints whatever the active backend gives it.

If the current model wasn't an exact id in the fetched list (an alias like `opus`, or a value set only in the shell), the picker fell back to the newest model, so pressing Enter changed the model. The current model now gets its own row and stays the default; keeping it leaves both CLAUDE_MODEL and CLAUDE_EFFORT untouched. Like CURSOR_MODEL, a value exported only in the shell counts as current. The Panel + numbered rows + Prompt.ask pattern repeated by the platform, backend, auth, model and effort pickers now lives in one helper, memclaw.prompts.choose. It sits outside setup.py so backends can import it at module level.

The fallback cost still used Sonnet prices, which is wrong once the model can be Opus or Haiku. Prices are now looked up per model (by id prefix, so dated snapshots match); an unknown model is estimated at the default model's price. The SDK normally reports total_cost_usd itself, so this only covers turns where it doesn't. Also renames _MODEL to _DEFAULT_MODEL to match cursor.py and refreshes the out-of-date "Sonnet 4 pricing" and "claude-agent-sdk 0.1.x" notes.

The env-isolation fixture and config helper were copied between test files, and the copy didn't clear the model/effort vars. They now live in tests/conftest.py as isolate_claude_env and the claude_config factory, used by the Claude backend, config and status tests. The model-discovery tests no longer need either, since fetch_models takes plain strings. Adds end-to-end tests for ClaudeAgentBackend.wizard_setup: auth choice, credential, model and effort, including a fetch failure that must still save the credential. The "(spec x.y)" references are gone, since that spec isn't in the repo.

The env-var table rows already cover the details.

memclaw imports mcp directly (mcp_tools.py, mcp_bridge.py) but never declared it, so CI's unpinned pip install pulled mcp 2.x, which removed the low-level Server decorators.

emanuilo · 2026-10-05T16:31:40Z

Its great now! Thanks a lot!

mistercap179 added 9 commits September 7, 2026 13:34

docs: note Claude model and effort selection in the README

ae01283

emanuilo self-requested a review September 23, 2026 01:33

emanuilo assigned mistercap179 Sep 23, 2026

emanuilo reviewed Sep 23, 2026

View reviewed changes

Comment thread memclaw/anthropic_models.py Outdated

emanuilo reviewed Sep 23, 2026

View reviewed changes

Comment thread memclaw/anthropic_models.py Outdated

emanuilo reviewed Sep 23, 2026

View reviewed changes

Comment thread memclaw/cli.py Outdated

emanuilo previously approved these changes Sep 23, 2026

View reviewed changes