Inference qualification roadmap#
Current implementation: Ollama and Strata are the two allowed backend values in Contract and ChatProvider. They supply inference. OpenHands supplies implementation tools and the agent loop. The three task lanes and their independent gates are documented in Workflow.
Preserve the working baseline#
Ollama advisory remains a supported path regardless of Strata's outcome. Qualified OpenHands/Ollama cases remain limited to their recorded identities, contexts, task scopes and resource profiles. A matching model name alone is not qualification. Fresh admission rechecks identity, resource fit and ownership. See Ollama/OpenHands evidence.
Strata: implemented, experimental, not enabled by this roadmap#
Strata is an alternative inference provider behind the existing OpenHands SDK → restricted mailbox relay → trusted OmniLAB worker → inference backend transport. The adapter and exact context accounting are implemented; historical patched text/protocol probes and offline SDK evidence do not establish live implementation qualification. See the existing implementation report and provider tests.
The next target is Qwen3.8-Flash-Next GSQ-RCO Coder, IQ1_M. Strata's official model guide recommends Coder for 32 GB systems. The linked artifact publisher is ISTA-DASLab, not the Qwen organization itself. Record that provenance accurately rather than calling an unverified conversion an official Qwen release.
At the 2026-10-17 inspection the personal Strata installation already contained
both target shards (29,608,446,496 and 28,800,138,432 bytes), and its frozen serving
configuration selected qwen3.8-flash-next-coder-iq1_m. No additional download is
needed merely to select this target. File presence and byte counts are inventory,
not proof of RAM fit. The addendum's full checksum reconciliation and pinned
publisher identities are recorded in the standalone guide.
Recheck all selected artifacts before the next authorized run. Do not silently
upgrade the pinned runtime.
The manufacturer's size recommendation is not this Windows machine's acceptance result: OS/apps, model mappings, GPU staging, cache/context, Docker and OpenHands must fit together with retained headroom. The existing memory, context, output, retry, ownership and cancellation gates remain unchanged.
Required sequence before implementation enablement#
| Gate | Required evidence | Current status |
|---|---|---|
| Provenance | Exact model shard hashes, publisher revision, engine/serving/tokenizer identity and license | Fresh shard/publisher and installed-artifact hash reconciliation passed; recheck before the authorized run |
| Preflight | Available/committed RAM, GPU use, pagefile activity, disk/evidence budget, idle competing workloads, exact LAB reservations and an exclusive owner | Must be sampled immediately before the authorized session |
| Standalone serving | Approved bounded localhost startup, browser/API checks, correctness, resource telemetry, cancellation, unload/shutdown and restart | BLOCKED: independent live launcher unqualified, approval consumed and resource gates fail; earlier narrow probes are historical |
| Real OpenHands work | Disposable fixture, exact permitted edits, actual file_editor and terminal invocations, protected baseline/candidate tests and independent diff review |
Not qualified for Strata |
| Negative cases | Forbidden path/test/tool refusal; invalid response; timeout; bounded retries; worker/provider failure and evidence retention | Offline coverage is distinct from live evidence |
| Cancellation/recovery | Owned process termination or honest unresolved-inference state, reservation retention/reconciliation, no replay, no effect on unrelated apps | Live qualification incomplete |
| Acceptance | Exact backend/identity/category/context/image/SDK record and independently accepted candidate | No production admission granted |
Use the existing selective contract and qualification fixtures. Require a baseline that detects the defect and protected tests that pass only after the permitted change. Retain tool messages, before/after hashes, test argv/exit status, output truncation, elapsed time, usage and resource peaks. An API response containing a tool call is not evidence that the tool executed.
Stop on admission failure, paging/driver/resource violations, ambiguous ownership or cancellation uncertainty. A proposal to change limits needs its own evidence and explicit authorization. A successful text answer does not waive these gates.
Independent Strata operation is a mandatory BLOCKED acceptance item. The installed Control opens, but it does not yet provide a qualified live Strata launcher. Browser/API source exists; current standalone inference, cancellation, shutdown and restart have not passed acceptance. Follow the separate standalone guide, then the OpenHands/Strata guide. Neither can be marked successful from installation or synthetic checks. Independent local-supervisor research is tracked separately in future directions.