feat: per-task LLM model/provider override without restart (fixes #129) by Uzzoper · Pull Request #155 · google/artemis

Uzzoper · 2026-09-26T03:31:23Z

Closes #129 with a small initial scope: a single optional override parameter, no settings UI, no global config mutation, no restart.

How it works

llm_model / llm_provider flow POST /api/run → enqueue_tasks (one shared normalize_llm_override() helper: trim/lowercase, blank → None, provider-requires-model fail-fast) → persisted queue item → worker --model/--provider flags → TaskRequestBuilder.with_llm_override() → per-task ArtemisContext → _resolve_endpoint precedence: task override > artemis.jsonc node > built-in default. get_llm() is the single funnel, so every model the task resolves is covered; the shared LLMConfig is never mutated.

Entry points

Console Prompt Dock: preset/provider dropdown fed by new GET /api/llm-options (providers, jsonc presets, configured default; degrades to free-text on 404) + custom free-text fields. Empty = server default; strictly per-task (never restored, cleared on successful submit).
Display: queue cards (override or scheduled default), history cards (model_info fallback), active dashboard, session detail — badges distinguish scheduled vs ran-with. Additive only, no restyle.
SDK: ArtemisClient.submit/run per-call params; daemon + batch shims (artemis run/batch --model/--provider).
MCP: mobile_run_task(llm_model, llm_provider) — named distinctly because existing model="Flash" means profile arch; background runner uses --llm-model/--llm-provider for the same reason.
Result: session row stores the keys in schemaless device_info (no migration); GET /api/sessions/{id} echoes them so TaskResult.llm_model/llm_provider resolves in production.

Also included (separate commit)

Fixes the 11 pre-existing ruff UP017/UP037/UP045 errors in playground/backend_manager keeping python-quality red on main — same mechanical fix as stalled #30/#67, needed for CI to go green here.

Validation (local, Python 3.12)

ruff format --check clean, ruff check clean, quality_ratchet.py at baseline (754/0/18), pyright --project pyright-core.json 0 errors
225 passed on touched backend suites (incl. new normalization/session-echo/isolation tests); full deterministic suite: pre-existing failures only (Flash credentials test_turn_index_snapshot fails offline when GOOGLE_API_KEY is not set #31/Flash unit tests require Gemini credentials before installing their model mocks #41, mock-provider _resolve_endpoint raises ValueError when provider/model in agent config is not a string #33, inspect_trace order flake test_mobile_inspect_trace_invalid_action passes or fails based on unrelated on-disk state #62, win32 flags test_ensure_emulator_uses_windows_creation_flags can only pass on Windows #97, adb tap docstring drift), verified identical with changes stashed
Frontend: ngc app+spec clean via --rootDir workaround (repo tsconfig lacks rootDir for installed TS — pre-existing; npm run build also gated on Node version locally, CI covers it)

Deliberate scope decisions (for reviewers)

CLI/shim/validation/fallback-pinning kept: without them the SDK/MCP params would be dead or fail mid-task; fallback pinning is documented, not incidental.
Provider-only is rejected (fail-fast 422/ValueError): model IDs are provider-namespaced, so provider-alone can never resolve.
No LlmOverride cross-signature type yet; kept the pair convention (verification_level/explorer_mode travel the same way).

Follow-ups (out of scope)

Client-side provider hint on provider-without-model; shared TS card component; --llm-model alias on artemis run.

google-cla · 2026-09-26T03:31:34Z

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

…d_manager Mechanical ruff check --fix plus format for the 11 errors keeping python-quality red on main (same as stalled google#30/google#67). Split out from the google#129 feature for scope discipline.

Implements google#129: optional llm_model/llm_provider override flowing POST /api/run -> task queue -> worker CLI -> per-task ArtemisContext -> _resolve_endpoint (precedence: task override > artemis.jsonc node > default). No global config mutation, no restart.

- Wire batch lane (--model/--provider through submit_batch_to_daemon and standalone run_batch_tasks, per-goal requests) - Fail fast on provider-without-model (builder ValueError, API 422, MCP pre-trace) instead of dying mid-task - Client: per-call override only (drop constructor defaults), provider lower-cased - Document fallback pinning; comment idempotent-retry setdefault - Tests: real-LLMConfig isolation, provider-only/invalid/Flash cases, run_task forwarding, batch override

Write llm_model/llm_provider into the schemaless device_info dict at session creation (worker side, run_tuning precedent; no migration), prefer them in model_service over the inferred/global model, and echo them from GET /api/sessions/{id} so TaskResult.llm_model resolves in production. Legacy rows fall back to current behavior.

Do not restore a previous override from localStorage: every task starts from the server default unless set for that task. Stale keys from earlier versions are dropped on load.

get_status() consults the session row override even when the worker set an active profile (single row fetch, no new queries); the session list prefers a stored override when no profile resolves. Rename profile_sess_row to sess_row_for_model.

Queue cards display the requested override (llm_model/llm_provider now flow through the status mapping); the session merge prefers the stored model_info over the global active model; a successful submit clears the override fields so the next task starts from defaults.

The second get_status() return path rebuilt model_info from the connection profile alone; it now reuses the session row (or one indexed lookup) and prefers a stored llm_model/llm_provider, with trace lookups still skipped on this path.

History loops get the same conditional model badge as queue cards; getTaskModelLabel falls back to the persisted session model_info when no explicit override keys are present.

Pending tasks without an override have no session row yet, so the badge had no data. getTaskModelLabel now chains explicit override -> session model_info -> global activeModel, showing the model the task will resolve at dispatch.

Single normalize_llm_override() helper (artemis/config/llm_override.py) replaces six copies; TaskResult echoes llm_provider; new GET /api/llm-options serves providers, presets and the configured default.

One shared getTaskModelLabel util; badges distinguish scheduled vs ran-with via title/aria; Prompt Dock gains a preset/provider dropdown backed by /api/llm-options with free-text fallback.

Rebuild the presets inventory test as a shape-only smoke test; fix the helper docstring coverage claim; stringify non-string override input in the client mirror.

Uzzoper added 13 commits September 26, 2026 13:45

fix: resolve pre-existing ruff UP017/UP037/UP045 in playground/backen…

aa2755a

…d_manager Mechanical ruff check --fix plus format for the 11 errors keeping python-quality red on main (same as stalled google#30/google#67). Split out from the google#129 feature for scope discipline.

fix: per-task-only override semantics in the Console dock

9afe1eb

Do not restore a previous override from localStorage: every task starts from the server default unless set for that task. Stale keys from earlier versions are dropped on load.

fix: show model on history cards with model_info fallback

eafacb1

History loops get the same conditional model badge as queue cards; getTaskModelLabel falls back to the persisted session model_info when no explicit override keys are present.

refactor: shared llm override normalizer plus options endpoint

64d96df

Single normalize_llm_override() helper (artemis/config/llm_override.py) replaces six copies; TaskResult echoes llm_provider; new GET /api/llm-options serves providers, presets and the configured default.

feat: shared label helper, badge distinction and preset dropdown

6130f05

One shared getTaskModelLabel util; badges distinguish scheduled vs ran-with via title/aria; Prompt Dock gains a preset/provider dropdown backed by /api/llm-options with free-text fallback.

fix: decouple presets test from workspace content, exact client mirror

58d8624

Rebuild the presets inventory test as a shape-only smoke test; fix the helper docstring coverage claim; stringify non-string override input in the client mirror.

Uzzoper force-pushed the feat/per-task-llm-override branch from 824560f to 58d8624 Compare September 26, 2026 17:15

chore: retrigger cla/google check

12457b3

This was referenced Sep 27, 2026

fix(playground): resolve ruff pyupgrade lint errors in backend_manager #30

Open

fix: resolve Ruff violations in playground/backend_manager #67

Open

Navigation Menu

feat: per-task LLM model/provider override without restart (fixes #129) - #155

feat: per-task LLM model/provider override without restart (fixes #129)#155
Uzzoper wants to merge 14 commits into
google:mainfrom
Uzzoper:feat/per-task-llm-override

Uzzoper commented Sep 26, 2026 •

edited

Loading

Uh oh!

google-cla Bot commented Sep 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Uzzoper commented Sep 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

How it works

Entry points

Also included (separate commit)

Validation (local, Python 3.12)

Deliberate scope decisions (for reviewers)

Follow-ups (out of scope)

Uh oh!

google-cla Bot commented Sep 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uzzoper commented Sep 26, 2026 •

edited

Loading