Skip to content

MiMo Code test samples

synthetic-* is synthetic data. real-0.1.15 extracts selected SQLite fields after the official client called a real local model. Eight parts match eight independent API usage records individually; provenance records the original database digest. Retain only placeholder identities, times, models, usage and required schema, excluding bodies, configuration and original paths. All eight finish reasons are length; CLI exit 0 does not establish task completion.

Source references (A14, fixed commit 456678b6a5afb0eef3fe2754575637218cfb3c84)

Current semantics follow official 0.1.15 commit 14dfe68a1c121f859544ba810b3c308e8501bfb2, session/session.ts:getUsage. It stores SDK input after subtracting cache and output after subtracting reasoning; adding those same components restores positive SDK totals. Default-zero subfields and costs stay unknown. OpenCode ancestry or a database’s highest version does not verify individual rows; use latest_fallback.

Scenarios

Directory Coverage
synthetic-step-finish Main/child sessions with three step-finish parts and a filtered text part; no cumulative columns, so no comparison
synthetic-unknown-version Unregistered session.version attempts latest_fallback
real-0.1.15 Eight CLI/session-continuation API/part calls; input 25,588, output 1,024, total 26,612, known cache read 20,733

Each directory’s _expectations.md contains manually calculated values. Unknown-format rejection and product separation (rejecting OpenCode databases) are constructed inline in tests/mimo_code_contract.rs. Default zeros, complete old summaries, consumed positions beyond samples, parallelism, rollback and conflict retention are covered by mimo_real_contract.rs. A single native request path does not verify every protocol, role or version.