Research

Code-context tooling for agentic development in 2026: the four paradigms, the current SOTA, our chosen stack.

Batch edits, fast-apply models, parallel search, and the skills ecosystem — the 2026 SOTA for faster agents that are also more reliable, plus the KPI schema we track them with.

Catalog of the checkpoint stack human developers built (lint, format, typecheck, git hooks, CI quality gates, security/supply-chain, mutation testing, spec-driven development) mapped onto the AI-native read→analyze→write loop: what transfers, what breaks, what is new. Includes the Wave 1.9 verify-stack adoption (mutation-score rung, format rung, per-repo verify commands), the ESLint-vs-Oxlint latency bench, and the verify-ordering pareto preregistration.

Living decision log for the memory stack: the adopt-now queue, the open cross-repo/freshness/fusion decisions with recommendations and empty result slots, and the rejected-with-evidence list.

Draft (v0.1). Proposes AgentDefinition — identity + capability contract + model set (0..n LLMs, the agentic gate) + versioned tool inventory + agents-of-agents composition DAG + memory/telemetry contract + runtime requirements + lifecycle state — as the load-bearing schema the @nexartis/cubicle-sdk epic compiler consumes to target local-mac, Cloudflare dynamic workers, or HB3 containers per declared needs. Names the six-stage MCP→agentic lifecycle (tool-prototype → tool-shipped → mcp-resident → kym-registered → composed → marketplace-published) with prior-art deltas against A2A AgentCard, NANDA AgentFacts, Anthropic Building Effective Agents, and LangGraph/AutoGen composition semantics.

One question in, one fused answer out: a single MCP that classifies, fans out to the relevant search experts in parallel, fuses ranked evidence with per-item provenance nuance, and emits a receipt of who fired. MoE applied to search tooling.

Draft v0.1. Extends ADR-007 with a new open-question intent class + a deterministic post-fusion synthesis stage: mine internal conventions (codesearch/serena/git-log-S + ADR/architecture-doc expert), gather external best practices (docs bundle + guarded web adapter), and emit a typed OptionPicker payload (2–5 options, per-option supports/dissent, single recommendation or honest undecidable, resolvable citations) with a signed receipt. Operationalises the AGENTS.md open-question rubric as a router capability rather than a prompt pattern.

Draft v0.1. Research-only evaluation of a code-graph expert candidate for the ADR-007 pool — LLM-free tree-sitter build of a typed, persistent code graph with EXTRACTED / INFERRED / AMBIGUOUS edge provenance, path queries, blast-radius traversal, god-node detection, ADR-as-node. Measured Graphify 0.9.29 (YC S26) on n=10 fixtures over nexartis-ozzydev + showrunner-link. Result: status-quo wins six, graphify uniquely answers three (god-nodes, provenance diagnostics, depth-3 blast-radius), one wash — the flagship path query is broken by high-degree stdlib hops. Recommendation: re-evaluate at a stated trigger (weight-aware path primitive OR ≥50-fixture bakeoff crossing 40% graph-wins). Nothing ships from this page.

The 2026 write-side of the memory stack: fast-apply models, atomic patch tools, and structural engines composed by a per-task routing rubric (a–h) — because a subagent that issues 20 sequential single-line edits is the failure mode we can actually fix.

v0.1. The ten Unix editing power primitives (ed addressing → sam structural regexps → patch algebra → mergiraf structural merge), which three LLM tooling has re-implemented, and the ranked gaps — first build track operator-locked 2026-08-01: mergiraf structural merge as a routed verb + git merge driver. The 2026 fast-apply landscape (commoditising; adapter, not moat), and the staged training program on the signed-receipt flywheel: Stage 0 data → learned router over the D8 rubric → apply fine-tune → hybrid verifier → north-star agent, each with kill criteria. D4: learned estimator OVER the rubric — splice stays primary; single LLM is the north star, not a 2026 commitment.

Living v0.1. How an orchestrator should oversee parallel coding agents that may work in the same repo — reliability (no lost work), context cost, parallelism ceiling. Five-level isolation taxonomy (serialise · shared-checkout+leases · worktrees · clones/VMs · speculative fan-out) with a decision matrix, the eight-clause dispatch contract distilled from C8 and the empty-report crash pattern, KPIs registered into the experimental-framework, and four open AGENTS proposals awaiting operator approval.

Landed (v0.3) — ADR-013 Phase 3. Ozzy Dev orchestrates harnesses; it does not marry one. Two protocol seams — ACP outward to engines, MCP inward to tools — admit every current and future agent that speaks them. The AGENTS.md §2 rule (a vendor name must never appear in a type union, config enum, registry key, adapter class name, or benchmark arm id) is enforced by an empirical test: a seam is not proven generic until a SECOND instance runs through it — and it now is. ONE AcpEngineAdapter landed on the UNCHANGED EngineAdapter seam, proving genericity at N=2 transport families (cli + acp); the wire is JSON-RPC 2.0 over newline-delimited stdio, NOT LSP Content-Length framing. Covers the headless AcpDriver contract (allow_once permission auto-answer, containment-only fs with loud path-escape rejection, distinct timeout/spawn/JSON-RPC/stopReason classification, honest cost/model tombstones), the instance ladder (acp.auggie LIVE-verified; acp.gemini-cli fixture-replay pending Google auth; Claude Agent queued), the ACP ecosystem inventory, the inversion pattern (vendor harnesses as MCP-clients of ozzydev-search/write, zero ozzydev code required), and the open conformance-target question.

Multi-armed-bandit study over agentic coding harnesses — same KYM agent cards + memory stack + skills, vary only the harness, score with signed receipts and dual (agentic + human) evaluations. First experiment: the three-arm retrieval-path bakeoff (auggie Context Engine vs auggie + our MCPs vs our full stack). Design v2 LOCKED (2026-07-29) at n = 800 tiered RH400/RM250/RE150, MDE δ = 0.10, 0pp falsification, internal-only pending pre-registration commit; Phase 0 BUILT.

Outline v0.1. Put a small, rigorously-defined agent (pinned LLM + versioned tool inventory + MCP-residency declaration) inside ozzydev-search and ozzydev-write, register it in KYM, and run dogfooding hypotheses as optimization-mode experiments over agent-definition versions. Five candidate skills ranked (write ambiguity-resolver first), the KYM schema/SDK gap list, and a three-phase plan where Phase A needs zero KYM changes.

Agents always have permission to write and execute scripts. Prior art (Anthropic PTC 85–98%, Cloudflare Code Mode 80%, Voyager 3.3×/15.3×, CodeAct +20%/−30%) justifies a curated skill library over ad-hoc scripts. Proposed nexartis-agent-scripts pnpm monorepo with a 5-step publish gate targeting WinterTC Minimum Common API.

The unified category × KPI × experiment-mode matrix nobody has published yet: six categories, the latency/accuracy/cost triad, five modes (including speculative-parallel N-worktree fan-out — a literature gap), one recording schema, plus concept registrations for the two ozzydev-native MCPs that operationalise it. Now carries the measurement discipline the bakeoff work made concrete: pre-registration hash-locked before the first metered run, a stated falsification criterion (a benchmark we cannot lose is not a benchmark) that is never post-hoc adjustable, sample size DERIVED from power analysis rather than chosen — including the trap that a cost-shaped n can masquerade as a statistical result, so re-derive when the constraint disappears — un-computable metrics returned as explicit not_observable rather than a fabricated zero, continuous [0,1] per-rung scoring over binary pass/fail, scorers verified against deliberately wrong input, and infrastructure crashes separated from genuine failures in any success-rate denominator.

Living log (v0.7). Migration 0002 landed the six D1 tables (experiments, variants, runs, turns, tool_benchmark_events, session_reports) plus the batch ingest routes and paper_slug/decision_slug write-back keys. Documents the schema, write paths, per-paper consumers, and the corpus roadmap — now including the workspaceId attribution wave and its durable lesson: a machine-global "active X" pointer is a misattribution vector for every fallback consumer, so when an artifact identity is knowable at emission time, EMIT it rather than letting a downstream guess. Covers the emitter-side resolveWorkspaceId fix across all three native MCPs and the bin CLIs (register-or-reanimate, loud fail-close, console unchanged), the decision to leave-and-annotate historical rows rather than mutate D1, the session_reports (session_id, agent_role, wave_slug) keying that silently collapses un-waved subagent rows, and the known rough edges.

Outline (v0.1). Long agentic sessions fill context windows; agents must self-monitor utilization + quality and report it. Traceability, compaction quality, handover-as-checkpointing, and the search-subagent (MoE-per-expert) pattern as structured-context injection. Test categorization: what the experimental framework measures locally vs what only the KYM receipt/eval architecture enables at scale.

Outline (v0.1). Voice-first sessions via the nexartis-remote-control speak channel: narrative updates, queued playback, ambient monitoring during parallel waves. Open UX questions on two-way control, interruption semantics, multi-orchestrator disambiguation, and DevFlow wizard step 2 integration.

The visual counterpart to voice-agentic-ux. Agents that reason, plan, and act need surfaces beyond a chat bubble: artifacts, generative UI, option-pickers, receipts dashboards. Corpus scan (Claude Code artifacts, Anthropic computer use, ChatGPT Canvas, Vercel AI SDK streamUI, AG-UI protocol, CopilotKit, assistant-ui, Thesys C1, Ink & Switch malleable software, CHI 2025 overview-detail) plus open questions and the Ozzy Dev integration surfaces (Kilo question/suggest option-picker, console /research + /memory + /roadmap, receipts + telemetry, DevFlow wizard).

Outline (v0.1). Composable agent + tool library deployed as Cloudflare dynamic workers, tuned in KYM, published to a Nexartis NANDA node for on-demand summoning from web apps. Runtime-capability categorization prefers dynamic workers over containers; containers via HB3 only when required.

Draft (v0.1). Solo-first, team-compounding: the same console is fully useful for one developer AND scales superlinearly with repos × teammates. The two-plane architecture (ADR-011) — shared team D1 for durable memory (secrets, receipts, telemetry, experiments, workspace inventory) vs per-workspace local tool planes (index, search, write, symbols) — is what makes it work. Prior art across monorepo tooling, meta-repo patterns, multi-repo AI orchestrators, CDEs, and team-memory MCPs; positioned as shared MEMORY + receipts, not shared compute.

First entry in a recurring News-format digest of the voices we track: per-voice 1–3 bullets with links, plus our own contributions this period (277/412/202 GB incident series, upstream issues, MAB-over-harnesses methodology).