Adopt

Adopt is the resumable discover → verify → recommend workflow with durable evidence. State lives in one JSON handoff file; every phase persists atomically and resumes where it stopped.

python3 scripts/mlx-agent adopt start --state ./adopt.json --role coding
python3 scripts/mlx-agent adopt resume --state ./adopt.json
python3 scripts/mlx-agent adopt status --state ./adopt.json

Verification evidence

Each shortlisted candidate gets an evidence record with an ordered strength:

Strength	Meaning
`runtime_measured`	Benched on this host (TTFT, tok/s).
`runtime_tested`	Generated and probed against a local runtime.
`runtime_inventory`	Present in a runtime inventory, not generated.
`metadata_only`	Hub metadata inspected; model not local.
`heuristic_only`	Name/metadata heuristics only.

Role probes

Deterministic, synthetic, never LLM-judged: coding-v1 (AST parse + sandboxed exec), reasoning-v1 (exact numeric answer), vision-v1 (OCR of a bundled synthetic image), embedding-v1 (cosine ordering), and tool-use-v1 (schema-valid synthetic tool call). A failed probe rejects the candidate for that role; an unsupported runtime is recorded, never penalized.

adopt start --measure adds a bounded measure phase between verify and compare that benches verified candidates and upgrades their evidence to runtime_measured.

Tool-use is canonical: metadata is never verification. Only a schema-valid synthetic runtime tool call recommends a model as tool-use capable.

Interpret and resume a result

The state records phases, shortlist, evidence, and decisions. Resume the same state after interruption; use status to inspect it without recreating the run. metadata-only status can include metadata or heuristic-only evidence, so inspect the strength separately. A recommendation does not establish that a candidate is installed or meets production workload requirements.

If no tool-use candidate is verified, make a shortlisted candidate available through the runtime or explicit Intake, then start a new adoption run. Adopt does not pull weights or start a server. Its bounded tool-use probe supports Ollama and local OpenAI-compatible LM Studio, mlx_lm, and LiteLLM servers. mlx-vlm reports unsupported-runtime for that role.

See Evidence and sizing before relying on a synthetic check, then use Wire to configure a reviewed model.