harness: GlideMQ evaluation run — research corpus, conditional-positive RFC, benchmark issue draft, research-rfc-run workflow

[Copilot]

Summary

Harnessed evaluation of glide-mq as a candidate queue backend / AI-orchestration layer for NetScript. Verdict is conditional-positive: GlideMQ hard-requires Valkey/Redis 7+ Functions + Streams (Garnet has neither; Deno KV/AMQP/Postgres structurally out), so it can only enter as an opt-in adapter behind the MessageQueue port plus NetScript-owned AI-execution port vocabulary — captured as an RFC draft with the required benchmark-prerequisite issue draft. Markdown-only; no packages//plugins/ source touched.

Scope

• Archetype / area: harness run artifacts (.llm/runs/), harness workflow docs (.llm/harness/workflow/)
• Related: informs epics epic: telemetry-revamp — framework telemetry convention + ports/adapters + dashboard query surface (Spine-1 enabler) #399 (telemetry harvest) and epic: NetScript Dev Dashboard — the Aspire/Scalar satellite that drives the framework (ships as a plugin, beta.6) #400 (dashboard prior art); refs epic: Road to 0.0.1-stable #301

Slices

• S1 Run activation — supervisor.md with lane table + blocked OpenHands evaluator lanes recorded
• S2 Research corpus + RFC — research.md (findings F1–F9, OQ1–OQ4), research/01–04 (seam baseline, GlideMQ dossier, compatibility matrix, ecosystem A/P/R mapping), rfc-glidemq.md
• S3 Issue draft + workflow promotion — issue-draft-benchmark.md, .llm/harness/workflow/research-rfc-run.md, activation.md step 11, close artifacts (plan/worklog/context-pack/drift)
• S4 Code-review nits (titles, clarity, heading)

Validation

• deno task check — n/a, markdown-only (no TS touched; verified via git status)
• deno task lint / deno task fmt:check — n/a, .llm/ prose is outside the package-quality fmt gate per netscript-tools
• deno task e2e:cli run scaffold.runtime --cleanup — n/a, no scaffold/plugin/db/aspire changes
• Secret scan on all 14 changed files — clean

Harness

• Run dir: .llm/runs/copilot-evaluate-proposal-and-documentation--glidemq-rfc/
• Phase: research → verdict/RFC draft. PLAN-EVAL pending — OpenHands lanes unavailable in this sandbox (blocked launch recorded in supervisor.md/drift.md); RFC and issue are drafts-only until owner ratification + separate-session PLAN-EVAL.

Key findings driving the verdict:

1. F1/F2 — backend exclusivity. All 44 GlideMQ server ops go through FCALL on a persistent Lua function library over Streams + XAUTOCLAIM. Garnet has no FUNCTION and no Streams → permanent blocker. Adapter shape only; portable floor (Deno KV/Garnet/AMQP/Postgres) preserved — precedent: sagas Garnet list-transport.
2. F4 — Deno risk. README claims Deno NAPI support; upstream HANDOVER.md admits it is untested. @glidemq/speedkey (personal Rust-NAPI valkey-glide fork) gates Track A behind a Phase-0 spike; WIRE_PROTOCOL.md FCALL specs are the pure-TS fallback.
3. F5 — bus factor 1, pre-1.0. Containment rule in RFC: no GlideMQ type on any NetScript public surface.
4. F6/F7 — durable value regardless of adapter. Its seven AI primitives (usage, token streaming, suspend/resume, failover, TPM, budgets, per-job locks) are the missing durable-execution layer of plugins/ai → proposed as backend-agnostic ports (Track B). Dashboard REST/SSE/authorize surface and OTel patterns harvest into epic: NetScript Dev Dashboard — the Aspire/Scalar satellite that drives the framework (ships as a plugin, beta.6) #400/epic: telemetry-revamp — framework telemetry convention + ports/adapters + dashboard query surface (Spine-1 enabler) #399 (Track C).
5. F8 — benchmark gap. packages/bench is the agent self-bench, not a transport benchmark; legacy netscript-start/benchmark is 404 → issue-draft-benchmark.md is a hard adapter prerequisite.

Also promotes the run shape as .llm/harness/workflow/research-rfc-run.md (stage contracts A–H, seed-run style), per the proposal's request to codify the missing workflow.

Drift / Debt

• Blocked OpenHands PLAN-EVAL/IMPL-EVAL launch (sandbox) — recorded, drafts-only fallback
• glidemq.dev DNS-blocked — substituted source-of-truth repo markdown, mapping recorded
• Legacy benchmark repo rickylabs/netscript-start 404 — issue scoped from first principles
• Tier-B research lanes ran on Copilot internal agents (not Opus 4.8); load-bearing claims re-verified first-hand
• Arch debt: none (no framework code)

Checklist

• Branch name and labels follow the taxonomy (exactly one status: label).
• Every referenced issue's acceptance + gate: boxes are checked with linked evidence (required before status:ready-merge).
• Docs/reference updated if public surface changed.
• Breaking changes are labelled breaking and, if substantial, backed by an RFC.
• No lock-file or unrelated churn committed (deno.lock only when a reviewed fix requires it).