Keystone Chat — Frontier Dev Stack (Research + Design)

0. TL;DR

One harness, two model profiles, selected automatically. The frontier of consumer dev is the harness, not the weights — every frontier tool (Claude Code, Cursor, Windsurf, Devin, Cline) is a harness over a cloud (or BYO) model, and theliterature shows up to 10× on coding benchmarks from harness/edit-format changes alone, no model change (Survey). That is the Σ₀ "the loop is everything" bet.

Keystone runs the same harness everywhere and lets the VRAM gate + serving-liveness probe (both already in local-model-registry.js) pick the model:

Same code path, same KEYSTONE_LOCAL_FIRST verify-gated escalation — only the threshold where local stops leading changes with hardware. No new serving path, no new memory system.

1. What "frontier consumer dev" means in mid-2026

The named discipline — harness engineering — has fixed components: tool/execution substrate, state persist + compaction, permission tiers, feedback routing, state-transition verification (Survey); context engineering = compaction + progressive disclosure + just-in-time retrieval. All map 1:1 onto the loop, so chasing the frontier is extension, not sprawl:

2. Model strategy — capability-gated ladder

Models stay interchangeable; this is a routing policy over the existing provider abstraction + VRAM gate.

Tier L-Frontier — LocalGB (leads when present)

Source: kilo.ai, InsiderLLM. Pick: Qwen 3.6-27B — dense beats the MoE on raw agentic quality (77.2 vs 73.4), and it matches Claude 4.5 Opus on Terminal-Bench 2.0 (59.3). When it verifies a task locally, you pay zero cloud tokens.

GPU note (Alex's box): ~17 GB Q4 fits a singleGB card — used RTX≈ $700–900, or rent (RunPod/Vast.ai/Lambda, aGB instance runs this comfortably; llama.cpp + speculative decoding measured 2.56× on this exact model/GPU).

Tier C — Cloud brain / ceiling (Claude)

Open-weight cloud backups (GLM-5.2, DeepSeek V4-Pro, Kimi K2.6) remain valid drop-ins, but Claude leads now per Alex.

Tier— LocalGB support layer (always, even on theGB box)

OnGB you load one model at a time; on theGB box this work co-exists with headroom. Cheap/frequent/private jobs that never need the frontier:

Every turn Tierabsorbs is a Claude token unpaid — a margin lever for a subscription/BYO-key product.

Why local can't always be the brain (honest)

A 7–9B model atGB tops out near HumanEval single-file work; its agentic SWE-bench is far below the bar — hence the gate. The 27B frontier model clears the bar but needsGB. The gate is exactly the line between "local leads" and "Claude leads."

Recurrent-depth (Ouro) — research-front

Ouro-1.4B is today's hardcoded Σ₀-native default (toolCalling:false). Genuine research front, but a 1.4B no-tools model can't anchor a tool-driven cockpit. Keep as a research registry entry; the default local lead becomes capability-gated (Qwen 3.6-27B ≥24 GB, else Qwen2.5-Coder-7B support).

3. Tech stack

3.1 The capability gate (the one new routing rule)

detect VRAM → selectAvailableChain(taskType)              # existing, VRAM-gated + liveness

  if best served local model.capabilityScore ≥ FRONTIER_THRESHOLD:

       local leads coding/reasoning  →  verify  →  on fail, escalate to Claude Opus   # KEYSTONE_LOCAL_FIRST

  else:

       local does Tier-0 support     →  Claude Sonnet leads agentic  →  Opus on hard

This is selectChain + capabilityFirst + the existing KEYSTONE_LOCAL_FIRST escalation, with (a) auto-detected VRAM instead of a hardcoded 8, and (b) the Qwen 3.6-27B entry registered with a frontier-tier capabilityScore.

4. The harness — the actual product (5 gaps vs Claude Code)

Already at parity (keep): verify-gated escalation (KEYSTONE_LOCAL_FIRST), permission tiers (command-allowlist.js + safe-exec.js), worktree isolation, PCSF routing, CSF memory, convergence records, web grounding.

The whole harness is GPU-independent and verifiable today against cloud Claude — the local frontier model just slots into the same loop when the GPU lands.

5. Roadmap

Recommended order: Phasenow (highest leverage, fully testable against Claude today, no waiting on hardware) + Phase 0a scaffolding in parallel (safe/inert onGB). Then 0b the moment the GPU arrives, then 2→4.

6. Open items for Alex

1. Cloud cost model: product fronts it (subscription, cf. Patreon tiers) or BYO-key? Sets how hard Tierworks.
2. GPU: buy (used RTX~$700–900 // 5090) or rent (RunPod/Vast/Lambda) — either works; the registry auto-detects.
3. Build scope: confirm Phasenow + 0a scaffold (recommended), or different order?

Sources

• SWE-bench Pro Leaderboard (Opus 4.8 leads, 69.2%)
• Best Local Coding Models by VRAM Tier— InsiderLLM
• Best Open-Weight Coding Models— kilo.ai
• Agent Harness Engineering: A Survey (10× from harness alone)
• Dive into Claude Code — arXiv 2604.14228
• Inside Claude Code's architecture — callsphere