[hongzuoj-pixel]

External CLI Agents previously required manually producing a patch and invoking verification. This adds mimo repair: a versioned argv/JSON adapter contract and a finite proposal → independent pytest acceptance → failure-feedback loop. Each proposal is cumulative against the original baseline; Agent workspace edits never become acceptance inputs, original source/tests are preserved, and only the last verified round determines the outcome.

• Separate Agent/acceptance workspaces and environment allowlists; bounded command time/output, credential redaction, process/container cleanup, strict completion/protocol errors.
• Durable invocation/time reservations shared across recovery branches, immutable per-attempt snapshots, matching-input proposal reuse and fresh tests on every resume. Opaque Agent token/cost usage is explicitly unavailable.
• Per-round patches and JSON/JUnit/log evidence; semantic history/accounting validation, portable export and PASS-only gates with final-patch/base binding; concise round/budget reports and Agent doctor preflight.
• Complete free two-round example, protocol/support boundaries, wheel/sdist inclusion and installed example coverage on Linux/Windows. Dedicated real-Docker repair acceptance. Version 0.9.0.

Validation before upload: installed-wheel full pytest 236 passed / 6 explicitly skipped Docker cases (pytest 9 / Python 3.12); 6 scoring invariants; independent installed-wheel + clean sdist example runs doctor → FAIL/PASS repair → resume with fresh tests → report → export → offline gate. Wheel's 16 package modules match source byte-for-byte. Wheel SHA256: 7ba1a95ab92bb5a2842727ae61ce1d474222ecfecdd157ef97df3a7036aa30ac.

Tested source tree: 7417126613cad48f2bd09605b65ac2f01294a7a4; preserved base: 2126f6901fed9327dc47e5decbd19a4efc2c165a. Require all 9 existing CI jobs, including real Docker, Python 3.10/3.12/3.13, pytest 8/9, Linux/Windows installed CLI and browser evidence checks, before merging. Vendor-specific paid CLI validation, additional Node/real defects, independent new-user trials and stable-release sign-off remain separate plan items.

PR validation: all 9 / 9 jobs succeeded on exact head f5d06c9301d5b7484dfcfe5c30b7b86ef09ef175: https://github.com/hongzuoj-pixel/MiMo-CodeHarness/actions/runs/39035207996 . Real Docker: 6 passed, including two-round repair and credential isolation. Windows installed pytest and the complete installed repair/resume/export/gate example also passed. Rechecked unchanged main base and mergeable clean state before merging.

Merged main validation: all 9 / 9 jobs succeeded on 59d3beb1b1577881508ef88476f55fbf8694fe4c: https://github.com/hongzuoj-pixel/MiMo-CodeHarness/actions/runs/39035509469 . Source tree matches the tested PR tree exactly. A separate focused follow-up tightens invalid UTF-8 response rejection before final handoff.