Skip to content

Add a portable ROMS harness over Decision Maker, knowledge and skills - #2

Merged
Buzburg merged 2 commits into
mainfrom
codex/decision-harness
Oct 7, 2026
Merged

Buzburg merged 2 commits into
mainfrom
codex/decision-harness

Conversation

@Buzburg

@Buzburg Buzburg commented Oct 7, 2026

Copy link
Copy Markdown
Owner

ROMS now has a portable preparation harness over its existing Decision Maker, indexed RAG/OKF, selected skills and MCP. A strict bounded request produces referenced context and a reviewable proposal or abstention through python roms.py harness and roms_prepare_harness. Preparation never executes or approves an action, calls a model or saves decision-discovery state. Existing Linux workshop approval remains separate; Omarchy is optional.

The review also fixed unsupported affirmative decisions for absent, irrelevant, negated and pending evidence, hardened numeric/input validation, and corrected public claims about model inference, performance and prompt improvement. README, GitHub-facing artwork and dashboard now describe ROMS in plain terms; legacy tool IDs and license attribution remain. Unmeasured legacy improvement/health metrics are now null with explicit status.

Validation:

  • Portable suite: Linux 138 passed; Windows 137 passed, 1 expected symlink skip; both also passed 24 subtests.
  • Existing release regressions plus final CLI/MCP checks: Linux 342 passed; Windows 328 passed, 14 platform/privilege skips.
  • Both real temporary-fixture demos pass. Harness tests cover read-only FTS/OKF retrieval, source/excerpt hashes, budget truncation, skill isolation, missing-evidence abstention and actual CLI/MCP transport.
  • Four-way GitHub CI matrix covers Windows/Ubuntu and Python 3.11/3.12.
  • Independent code and documentation review completed; original dirty ROMS checkout is untouched by this change.

Limits: lexical scores remain advisory, indexed source files are not revalidated, and SQLite can use coordination sidecars. macOS/native-model/GPU performance is unverified. The optional legacy embedding-model integration test could not run offline because local all-MiniLM-L6-v2 weights were incomplete; targeted prompt/inventory tests passed. No model-weight or OS installation is part of this PR.

@Buzburg
Buzburg merged commit 2d2e207 into main Oct 7, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant