Repository navigation
[LLM-Wiki-v2][F2] Optional upstream engine adapters: candidate-only parity pilot and fallback #875
Description
Logical parent: #871. Depends on: #873 (plain issue links; no native GitHub subissue claim).
Scope and acceptance
Only if Arena recommends: test actual pinned Atomicstrata SDK compile and/or Geronimo binary search/graph on an isolated fixture, not a mocked success path. External engine must have no direct PStack/QStack canonical mutation authority. Emit bounded candidates with source revision, content hash, model/prompt identity, provenance, review status; use QStack existing publisher on QStack, never upstream review approval. Verify no provider => BLOCKED/typed failure, offline baseline remains functional, optional binary absence leaves a documented gap, no leaked secrets/network or clobber. Measure equivalence, headless performance and rollback; do not adopt just from README claims.
Evidence policy
Acceptance requires a reviewable PR and exact-head reproducible receipts. Planning is not implementation; no claimed Swarm/Arena/Interrogate model execution without actual independent work.
Activity
tommy-ca commentedon Oct 31, 2026
Resolved as not_planned per Arena decision #873 and PR #880 (docs/research/llm-wiki-arena-decision-2026-10-31.md). External reference engine adapters (Atomicstrata Node.js compiler, Geronimo Rust engine) are deferred to preserve zero-dependency offline CLI portability across all six supported hosts without adding speculative dual-engine complexity. Option A (pure Python stdlib) is the canonical baseline.