Atrex Kernel Agent v0.3.0 — Long-horizon optimization and production campaigns

Atrex Kernel Agent v0.3.0

This release covers changes since v0.2.0 (July 18, 2026). AKA now runs GPU kernel campaigns through one orchestrated entry point, with longer-lived optimization episodes, broader evaluator and hardware support, and stronger supervision of correctness, recovery, and promotion.

Highlights

• One optimization workflow. orchestrator/optimize.py is the supported entry point. The internal Long Horizon engine explores each version in an isolated Git worktree, records a structured journal, and promotes only a verified improvement. The first two episodes use five fast evaluator-backed trials by default; later episodes use profiling and same-allocation incumbent/candidate ABBA verification. Interrupted episodes resume in their existing worktrees. After sustained stalls, campaigns can schedule a broader goal episode while retaining the normal promotion gates. (#38, #49, #63, #66, #85)
• Production campaign control. AKA accepts both SOL-ExecBench and native Atrex-Bench operators. Omitting --framework dispatches independent campaigns for the frameworks supported by the detected GPU. Production mode adds a framework-native V1 baseline, private detailed-shape handling through a public agent_problem.json contract, independent dependency and framework review, and final policy checks. Triton campaigns can escalate to Gluon after repeated stalls. (#34, #36, #37, #52, #55)
• More coding-agent and reviewer choices. Claude, Qoder, Codex, and Pi share one Agent Runtime interface. Codex supports resumable episode threads; optional Codex and Qoder plan reviews can be configured separately for V1, fast episodes, and full episodes. Per-phase and per-episode token telemetry improves budget and cost visibility. (#35, #40, #41, #43, #46, #58, #70)
• Isolated GPU execution and recovery. Campaigns can use a local sandbox or GPU gateway, or run on a directly connected OpenSSH host inside a networkless Bubblewrap namespace with one assigned physical NVIDIA GPU. Confirmed environment outages preserve the candidate and start durable, supervised recovery. The sandbox also supports two-GPU distributed evaluator requests (torchrun/NCCL), OSS-backed output downloads, and a community local gateway. (#33, #51, #77, #79, #80)
• A structured GPU knowledge flywheel. GPU Wiki now separates ranked kernel experience from exact hardware facts, supports architecture-scoped and natural-language retrieval, mines optimization and session traces, and records which retrieved knowledge influenced an episode. A new repository-local plugin interface discovers tools, Skills, instructions, and resources automatically; GPU Wiki is the first bundled plugin. (#27, #59, #60, #65, #72, #84, #87)
• Expanded profiling. New timeline tooling helps instrument CUDA and CuteDSL kernels. T-Head PPU support adds ACU profiling, adaptive PPU timeline analysis, and optional correlation between the two; PPU reference projects are included. TileLang production campaigns are supported. (#69, #73, #74, #76)

Correctness and measurement

• Full-workload evaluator checks remain the basis for candidate acceptance. V1 and full episodes add multi-seed checks; full episodes also use correctness-passing ABBA verification. Native FP4 operators use relative L2 error with a 0.2 threshold; other native operators keep their ordinary evaluator comparison policy. Performance scoring now uses the arithmetic mean of per-shape speedups consistently across evaluation and promotion. (#67, #95)
• Numerical review can propose bounded supplemental input distributions. The supervisor runs applicable probes through the immutable evaluator, sends measured candidate failures back to the coding agent for repair, and returns invalid reference inputs to the planner. Probe plans and coverage receipts survive restarts; the supervisor owns the supplemental error thresholds. These probes add validation work but do not replace ordinary correctness, policy, or performance gates. (#96)
• Promotion is tied to the reviewed kernel and solution manifest. Failed correctness, policy, or performance checks retain evidence without replacing the incumbent. (#63, #96)

Upgrading from v0.2.0

• Launch campaigns with python orchestrator/optimize.py; the former interactive SKILL.md route and install.sh have been removed. See Quick Start for current commands and options.
• Pass --framework when you want one campaign. Without it, AKA starts one campaign per supported framework, each with its own workspace and budgets.
• Production and leaderboard campaigns use separate workspaces. Start a fresh production workspace if an older run exposed detailed evaluator shapes to the agent.
• Expect different performance scores from v0.2.0 because the aggregation method changed. Compare kernels using the current evaluator and promotion path.
• OpenSSH mode requires an explicit framework and a physical NVIDIA GPU index; MIG assignment is not supported. Distributed evaluator requests currently support the validated single-node, two-GPU contract.

For the current CLI and architecture, see the Quick Start, Architecture Design, GPU Wiki, and Plugin Guide.