MiMo, Zoo and DSH native-source and artifact acceptance
2026-10-06, version 0.2.1; implementation prerequisites and three-source field rules. Official clients called a real local model in isolated rootless Podman. Native files, model API and application SQLite were compared independently. No personal accounts/paid cloud requests were used. Runtime had no external network/host mounts; ordinary UID 1000, zero effective capabilities, default seccomp. Downloads/image preparation were separate from offline runs; installation/empty sessions do not verify usage.
Fixed distributions and actual source
| Product | Official reference | Actual distribution |
|---|---|---|
| Zoo Code 3.86.0 | Release，commit 6aa9d0174a9ecae155c6c5db9134bead4b67197d |
VSIX 34,635,449 bytes，SHA256 25c338c866bf7dacde840d8039093de867c2b0f70070d68d50031a04fe3b5f73 |
| MiMo Code 0.1.15 | Release，commit 14dfe68a1c121f859544ba810b3c308e8501bfb2 |
tar 46,617,071 bytes，SHA256 3530927a2d69eb0f809c11f663ecd6939b82c0b598ff0fbe1f82f431d764ca0c；ELF 134,998,144 bytes，SHA256 872728c1547b9910bb8d98b5d2fe9b1a461649b6a3e147232188f020f0ddc2e8 |
| DSH 0.2.0-rc.2 | Official npm manifest, installed lockfile and compiled modules | SHA1 dfc8f7e09cfa96b854d6f0cf3a973ce7f2948925; SHA512 integrity below; current GitHub commits do not verify rc.2 |
DSH integrity: sha512-EAJ3gPNcVt/uv8X19PMm9NkVhWgT7xXNMk0UKCVm+IQ5rpSQOcsMUa0HWlnYYVybKMsccjcRB21vVVsaXQ6IdA==. Checked installed session-format, JSONL persistence, LLM codec, token-meter and pi-ai 0.87.1 individually; current GitHub code does not establish npm package behavior. Zoo checks covered the full message enumeration/provider defaults; MiMo checks covered that release’s session getUsage, Global/DB paths and SDK 6 normalization.
The official Qwen3.5-0.8B GGUF BF16 model is 1,557,662,496 bytes, SHA-256 9a7bed4041b7975e0f71fa34670d1e9025213bc92905ac0db75d36c4fa3fa623. Local llama-server used context 64,000; a transparent proxy forwarded actual SSE/usage unchanged without constructing responses/tokens. This sample-generation environment does not establish hardware performance.
Native samples and coverage
| Source | Native/API comparison | Limits |
|---|---|---|
| Zoo | One primary call: input 6,118/output 53/derived total 6,171; five native UI messages, public extension usage callback and API matched | Actual VS Code 1.140.0 GUI/public API; cancelled after receiving usage, task completion unverified. Default-zero cache/cost unknown; uncached input/model/provider/source total unknown. CLI/deletion/compaction need separate checks |
| MiMo | Initial/resumed session: eight step-finish records/eight API calls; input 25,588/output 1,024/total 26,612; positive uncached input 4,855/cache read 20,733 | Both CLI runs exited 0; all eight finishes were length, so task success is unverified. Read part only, without adding message usage; zero cache write/reasoning/cost unknown; each owning session’s version remains latest_fallback |
| DSH | Headless/resumed runs completed; two native settlements matched main-loop API: uncached 5,606/cache read 5,568/total input 11,174/output 105/total 11,279 | Another title API call used 205 tokens but no native result was saved, so none is invented. Stream usage is a duplicate; native computed total is not source_total. Other protocols and native seed/retry/fail scenarios unverified |
Original roots are under the independent WSL workspace’s build/plan-final-push/: Zoo zoo-real-1791290947, MiMo mimo-real-1791291189, DSH dsh-real-1791291190. Zoo UI: 986 bytes, SHA-256 402e3148ae2242dfa0840193071fc922c78bf34316c6324c779420dc95bd5fec. MiMo main DB: 401,408 bytes, SHA-256 0a60be4932cb9b3cd543479a713885ca9ad9045fc16991f488b26681299641cd; WAL: 1,516,192 bytes, SHA-256 ca0343b73d1c1e281ed06fdd321ca046aad640bd8e4cce6c58b43478f53726df. The last four MiMo records were still in WAL; the main DB alone does not verify all eight. DSH zstd: 17,575 bytes, SHA-256 bfe960a2d99eb2726c08d839a261dbe3ce645801e10b69af205e07159d20fee7; decompressed: 51,114 bytes, one v4 header/27 consecutive-seq events.
Test data under core/tests/fixtures: zoo/real-3.86.0, mimo-code/real-0.1.15 and dsh/real-0.2.0-rc.2. Only per-type allowlisted fields remain, with anonymized identities; bodies/requests/configuration/ private paths/credentials/original DBs are not committed. DSH redacted records preserve event types/order/time/usage and reduce other payloads; they do not claim complete upstream restoration. New/prior sample audit: 68 files/171 JSON objects, zero credential keys/private paths.
Corrections and targeted regressions
Zoo supports actual ask/say enumerations, rejecting the whole file for unknown enumerations; default-zero corrections match complete old event summaries. MiMo uses product-specific normalization/environment paths and protects version ownership/quality/model/real conflicts. DSH adds bounded independent v4 JSONL/zstd reading under token-meter final settlement/retry/ inheritance boundaries. Future generation versions are rejected without falling back to old generations; partial writes/limits retain cursors. Default/manual-root/physical-file deduplication was checked through the full registry. Complete native headers may restore old incorrect ownership only without statistical history; owned files/checkpoints transfer transactionally, while other invalid files/diagnostics remain. Kimi Work’s observed absolute root applies only to the current process home; other-user test contexts do not probe it. Native tests must still isolate the actual child process’s USERPROFILE/HOME/source environment.
New Zoo five/MiMo five/DSH v4 nine requirement tests passed, alongside legacy Zoo five/MiMo five/DSH four and OpenCode/Roo/Kimi Work/routing regressions. Cases cover complete canonical/ legacy old summaries, consumed cursors/processing positions, repeat reads, concurrency, transaction rollback, same-batch conflicts, protected fields, missing/invalid buckets, inheritance/retries, snapshot decreases, zstd window/decompression limits and unknown events. Constructed boundary cases are explicitly synthetic regressions.
Artifact acceptance for this batch
Commands/exit codes/failure logs remain under build/plan-final-push, including historical first failures.
- Windows npm run verify exited 0: 1,007 Rust tests, core 906/app 101, eight ignored by default; 21 frontend/four script tests; Markdown 201 files at that stage. Svelte had no errors/warnings; fmt, Clippy -D warnings, assets/frontend build passed. Later documentation changes were checked separately.
- Debian: 1,004 Rust tests, core 906/app 98, nine ignored by default; native Clippy and release/ deb/AppImage builds exited 0.
- Windows NSIS 12/headless 11/WebView2-IPC 17 checks with 20 startups/receiver eight passed. Initial-page P95 751.36 ms; no owned credentials/installation integrations remained.
- Linux root 1791295818331: nine groups/47 checks covering deb install/upgrade/rollback/ uninstall/reinstall/purge and FUSE/GTK/Orca, exit 0. Actual focus/speech were checked across five pages/ten languages. Only the owned container received SYS_ADMIN; application-user capabilities remained zero with default seccomp. Actual FUSE mounts/release on exit were checked. Host login/logout, complete assistive technology, language pronunciation and physical audio remain unverified.
- Final Markdown 202 files; 44 affected Markdown files/262 local references; fmt, Skill quick_validate and git diff –check exited 0. Skill checks explicitly used Python UTF-8. Read-only cleanup found no Podman containers; an old owned, never-started Roo reread container was removed.
| Artifact | Bytes | SHA-256 |
|---|---|---|
| Windows exe | 10,106,880 | faec1e22d20756c41a301bc5727e68ed9b7325bd53d76ff3a183ecab2169b56c |
| Windows NSIS | 3,985,629 | cba1db1a17432b3ab56625f3d55d7b5b692b303c685ee8139e2a7e9b1cbaa7ab |
| Linux deb | 5,527,382 | 4afb1bd7958983de49a121966055a297f5ee8e50432e70b3c6511db71bd22854 |
| Linux AppImage | 111,204,856 | c08b20ad16569ce1225360426ef515b6fc7c1bbfb5ee78f1e95ffbb01501bc99 |
Linux m3-readback-1791295919 and Windows m3-windows-1791296331720 created old databases with the previous OpenClaw-stage package, then upgraded and repeatedly read unchanged native sources. Old package: Zoo zero/MiMo eight/DSH zero; new: one/eight/two. Three physical files/checkpoints, zero conflicts, health ok. Old event identity/revision/first observation and two prior audit records remained. Repeated scans added/changed no committed events. Main DB/WAL/JSON/zstd hashes remained unchanged; SHM bookkeeping was excluded independently. Windows verifies application reading of the Linux client files, without verifying official Windows clients.
The same package reread 15 prior sources: Aider, AtomCode, Continue, Crush, gajae-code, Goose, Hermes, jcode, Junie, OpenCode, Qwen, Roo, Xum, Cline and OpenClaw. All exited 0; summary logs: m3-prior-readbacks-1791296018. Values, unknown components, old-database upgrades and existing coverage gaps remained unchanged.
First failures and corrections
- DSH development failures included duplicate entry/file probe APIs, TokenQuality lacking Eq, wrong SQL table names/expected counts. Logs remain; fixes use actual interfaces. Full checks found Option::is_none_or newer than MSRV 1.77.2; compatible code passed without raising MSRV.
- Full-registry discovery initially let Pi claim a manual DSH type=session file. Recovery now requires a complete native fingerprint and no statistical history; protected-history/other-file tests passed.
- Initial Linux reread copied MiMo main DB without WAL, so only four records were readable; SQLite’s new empty WAL also caused the hash assertion to fail. Copying the complete source produced eight records matching API. A separate temporary-script assertion used dsh instead of deepseek-harness; that failure remains recorded.
- Initial Windows scripts broadcast manual roots, producing legacy Cline/other-product misidentification diagnostics; invalid TMP caused SQLite commit failure. Independent single-variable checks showed exit 1 with absent TMP and exit 0/eight committed records after creating it. Scripts now use verified product-specific isolated source paths and valid temporary directories. A later omitted repository isolation helper let libuv add USERPROFILE and scan other default sources. That result was invalidated, its extracts not committed; corrected startup isolation discovered only the three test sources.
- The fourth native GUI startup chose port 6669. Node fetch rejected it before connecting, producing a CDP timeout. Local fetch reproduced bad port, and the Fetch standard blocks it. The shared HTTP-port probe now completes a real local 204 request through the same fetch, retries rejected ports within bounds and closes its owned server before CDP use. Twenty probes, seven script syntax checks and 20 actual GUI startups passed. Later passes retain the first-failure explanation.
No macOS desktop/specific-hardware acceptance occurred. Other clients/native mixed versions/ protocols/host and remote-CI conditions remain in Plan.md; this batch does not complete every milestone. On 2026-10-06, a read-only check confirmed baseline CI 37330787284 had all eight jobs successful for HEAD c2c8f6f on 2026-10-05, with three-platform artifacts. This batch’s uncommitted changes are absent from that revision. Official repository references were checked for all five current Action tags/branches. The user then authorized an independent test branch, commits/pushes and three-platform CI; results are in this batch’s CI record.