IMPROVE THE WORKFLOW · KEEP THE HUMAN GATE
Better through evidence.
Not automatic self-rewriting.
MOMM can review changes to MOMM. That does not give a reviewer permission to change its instructions, publish private work or release software.
The improvement loop
- Observe: a release check records public version and asset metadata, tied to an exact commit.
- Suggest: observations go to GitHub Issues under a release-specific MOMM title. Unchanged reruns do not add comments.
- Investigate: an engineer reproduces a suspected defect and writes a regression test. A model's agreement is not proof.
- Decide: the governor records applied, rejected or deferred, with a reason and evidence.
- Review and release: an isolated candidate passes tests and peer review; a human authorizes publication.
Release observations on GitHub → Report an observation →
The repository workflow runs on each stable MOMM release and on a maintainer's manual request; it has run on the default branch since 27 September 2026 (for example the 1.16.1 run). It posts bounded public metadata only. This page is not proof that the workflow has run; follow the linked issue and Actions run.
What never happens automatically
No model calls, capability re-probes, private ledger uploads, code fixes, merges, instruction changes, skill updates or releases. The observer does not mark a release READY. Closed observation issues stay closed; changed observations are bounded to one summary update, not repeated model suggestions.
Released work and follow-up proposals
1.16.1: released on 28 September 2026 as the signed tag momm-1.16.1. It closed proof and safety gaps, including installation inventory, source-bound receipts and clear dirty-clone recovery. See the 1.16.1 release notes.
1.17.0: released on 3 October 2026 as the signed tag momm-1.17.0. It adds safer reviewer isolation, source-bound evidence and stricter completion checks. Read the 1.17.0 release notes for shipped changes and testing limits.
Unshipped ideas remain follow-up proposals, not promised features or an approved next-release scope. The historical ideas register records earlier proposals, not the current release status.
Can a reviewer actually inspect your attachment?
Use the shipped adapter limits, current capability evidence, the actual workload and the runtime result together. Vendor branding, model names and extensions do not prove perception. Local preflight makes zero model calls; active synthetic media probes are a separate, consented step. An expired or failed capability check must not count toward a media quorum.
A text-only reviewer cannot dismiss a visual finding by saying the source looks fine. Investigate the visual claim with the matching evidence; do not turn either model's opinion into an automatic blocker or approval.
Send a useful observation
Give the version and full commit, the smallest non-private reproduction, expected and actual behavior, and the relevant trimmed output. Remove tokens, account identifiers and local paths. Do not upload your private review ledger. For a security-sensitive report, use the repository's private reporting route if available; do not disclose an exploit publicly.