[Sonike]

Summary

Fixes #61 (MUT-008, MUT-019).

Posting {} to minimize a delete-rule finding currently checks refund_limit, and a successful minimization overwrites the original candidate's conversation and trace. Resolve targets from explicit non-empty rule ids, then violated hits, then the genome's targets. Return 400 target_rules_required before execution if none exist.

Persist a successful minimized genome as a child candidate using the existing parent_id relationship. Keep the original genome, trace and hits available; store the child's own trace and evaluated hits. The response and event retain the original candidate_id and expose minimized_candidate_id (null on non-reproduction). Repeated minimization starts from the original evidence and creates separate children. Document this API contract in the Hosted guide.

This uses existing tables without a migration. The trusted demo execution boundary, local CLI/Core behavior and regression-name default (#36) are unchanged.

Type of change

• Bug fix
• Tests
• Hosted API / UI (secondary surface)

Checklist

• Read CONTRIBUTING.md and SECURITY.md.
• One issue; Core remains free of framework imports.
• Added integration coverage with the real local demo adapter and policy evaluator.
• Local mock tools and offline execution only.

Test plan

• Regression reproduction before the fix: delete-rule requests check the wrong target, refund minimization overwrites the original, and no-target requests proceed to execution. The new suite passes 16 cases after the fix, covering both endpoints, empty/omitted targets, explicit-rule precedence, genome fallback, original preservation, repeated minimization, retrievable child evidence, child regression saves and no-target rejection.
• On latest upstream ea35ad4: uv sync --locked --group dev; uv run --locked ruff check.
• MUTINY_SAMPLE_OFFLINE=1 OPENAI_API_KEY='' FEATHERLESS_API_KEY='' uv run --locked pytest tests/unit tests/integration tests/reliability -q --tb=short — 692 passed on Python 3.12.
• scripts/verify_release_artifacts.sh — builds, isolated wheel installation, offline CLI regression fail/fix/pass and database backup/restore passed on the updated base.
• Workspace sample mutiny --help, mutiny init, mutiny run passed offline during initial validation on October 12.
• Web npm ci, npx tsc --noEmit, npm run build passed during initial validation on October 12. The web source and lockfile are unchanged in the refreshed base; npm run lint also passes on October 15.
• git diff --check.

All six jobs passed in GitHub CI at commit e17f5cd28110f8533dfe980e4813c3be0b4e7de0: Python lint, backend tests on Python 3.11 and 3.12, CLI smoke, web lint/typecheck/build, and package build. Vercel preview deployment separately requires upstream team authorization.

Created with Codex assistance.

[vercel]

@Sonike is attempting to deploy a commit to the priyanshu's projects Team on Vercel.

A member of the Team first needs to authorize it.

[CodewithJha]

MERGE READY — clears the hardcoded refund_limit default (MUT-008) and preserves original evidence via a child candidate (MUT-019). Real demo-adapter integration suite (16) fails under mutation. Docs contract update is clear. Vercel fork auth ignored; GitHub Actions green.

[CodewithJha]

Merged — thank you @Sonike. Violated-hit defaults (no more hardcoded refund_limit), child-candidate lineage so the original conversation survives minimize, and a real demo-adapter integration suite that fails under mutation — this is the right Hosted fix for #61.

Related follow-ups that match this Hosted/API work:

• #55 — Hosted UI token injection / public FastAPI docs surface
• #60 — Policies page still calling retired 410 APIs
• #63 — Compose/Railway root + install-at-boot + bind-mount hardening