Add GEAK v4 agent integration

Makefile

@@ -6,7 +6,7 @@
 
 SHELL := /bin/bash
 
-.PHONY: help docker-shell docker-check-agents docker-smoke docker-run docker-parallel-run docker-setup-flydsl \
+.PHONY: help docker-shell docker-check-agents docker-smoke docker-run docker-parallel-run docker-setup-flydsl docker-setup-geak \
         check-docker-runner check-evaluator check-held-out check-visualization \
         visualization-build visualization-serve visualization-run \
         sync-perf-helpers check-perf-helpers materialize-perf-workspace \
@@ -27,6 +27,7 @@ help:
 	@echo "                         On other GPUs, pass a matching CONFIG explicitly"
 	@echo "                         Images: gfx942->mi30x, gfx950->mi35x; override with AKA_DOCKER_IMAGE=..."
 	@echo "make docker-setup-flydsl - Install FlyDSL when absent (for flydsl2flydsl, torch2flydsl, and triton2flydsl)"
+	@echo "make docker-setup-geak   - Install the Claude Agent SDK when absent (for the geak_v4 agent)"
 	@echo "make check-docker-runner - Check Docker runner syntax and runtime-specific arguments"
 	@echo "make check-evaluator     - Run centralized evaluator unit tests"
 	@echo "make check-held-out      - Run held-out module unit tests"
@@ -74,6 +75,11 @@ docker-parallel-run:
 docker-setup-flydsl:
 	@$(DOCKER_RUNNER) setup-flydsl
 
+# Install the Claude Agent SDK into the container's persistent pip user-base when
+# the selected image does not ship it. Needed by the geak_v4 agent.
+docker-setup-geak:
+	@$(DOCKER_RUNNER) setup-geak
+
 check-docker-runner:
 	@bash tests/test_docker_benchmark.sh
 

agents/geak_v4/README.md

@@ -0,0 +1,108 @@
+# GEAK v4 Agent
+
+The `geak_v4` integration runs GEAK's deterministic
+`kernel_workflow/kernel_workflow.js` through Claude Code's dynamic **Workflow**
+tool. GEAK optimizes the kernel and, when its Director validation passes, applies
+the validated patch directly into the task workspace (`apply_to_original="true"`).
+
+**AgentKernelArena is the single source of truth for scoring.** After GEAK
+finishes, AKA verifies its harness is intact, re-materializes the perf helpers,
+and independently re-evaluates the kernel (compile → correctness → performance);
+GEAK's own numbers are not used for the final result.
+
+Like the `forge` and `claude_code` agents, `geak_v4` relies purely on the
+environment — it touches nothing outside `agents/geak_v4/`. The launcher stays
+thin: it writes a versioned handoff and shells out to `workflow_runner.py`.
+
+## Why a runner instead of a plain `claude -p`
+
+On current Claude builds the `Workflow` tool runs as a **background task**: the
+main agent turn ends immediately, so a one-shot `claude -p` would return (and
+tear the still-running workflow down) before GEAK finishes. `workflow_runner.py`
+keeps a persistent `claude_agent_sdk` client alive and drives completion off the
+SDK's background-task lifecycle (`TaskStartedMessage` → `TaskNotificationMessage`)
+plus GEAK's on-disk terminal marker. It is a kernel-scoped analogue of GEAK's own
+`interface/run_e2e.py` and mirrors its `handoff.json` / `result.json` contract.
+
+## Prerequisites
+
+- An AMD Instinct GPU and a supported profiler (`rocprof-compute`).
+- A local GEAK checkout. Point `GEAK_V4_WORKFLOW_DIR` at its
+  `kernel_workflow/` directory (default: `/opt/geak/kernel_workflow`).
+- Claude Code 2.1.177 or newer, installed and logged in (the minimum version for
+  the dynamic Workflow feature), plus access to the model configured in
+  `agents/geak_v4/agent_config.yaml`.
+- The `claude-agent-sdk` Python package installed in the interpreter that runs
+  the agent.
+- For a `flydsl2flydsl` task, FlyDSL available in the environment.
+
+## Setup
+
+```bash
+# 1. Clone GEAK and expose its kernel_workflow directory.
+git clone https://github.com/AMD-AGI/GEAK.git
+export GEAK_V4_WORKFLOW_DIR=/absolute/path/to/GEAK/kernel_workflow
+
+# 2. Authenticate Claude Code and confirm the Workflow-capable version.
+claude --version
+claude            # log in
+claude auth status
+
+# 3. Install the SDK the runner uses to hold the background Workflow open.
+pip install claude-agent-sdk
+```
+
+This adapter was developed against GEAK commit
+`4965d5b2ccde927925c8c5501a25c1233daa52eb` (`v4.0.0-102-g4965d5b`). For a
+reproducible review, check out that revision; newer GEAK revisions may require an
+adapter/schema update.
+
+## Run the example
+
+The included MI300/MI300X example runs one HIP GELU task. Run it in an
+environment that already satisfies the prerequisites above (`GEAK_V4_WORKFLOW_DIR`
+set, `claude-agent-sdk` installed, `claude` logged in):
+
+```bash
+python main.py --config_name example_configs/quickstart_geak_v4_mi300.yaml
+```
+
+## Supported task types
+
+- `hip2hip`
+- `triton2triton`
+- `flydsl2flydsl`
+
+These are single standalone kernels. The launcher reads `source_file_path` to
+steer the optimizer ("optimize only these files") and fails early if the declared
+anchor source is missing. Authoring, translation, repository, and image-level
+tasks are out of scope for this integration.
+
+## Artifacts
+
+For a task workspace named `<workspace>`, GEAK's run artifacts live OUTSIDE the
+scored workspace, under a hidden sibling directory:
+
+```text
+.<workspace>_geak_v4/<run-id>/
+├── eval/           # GEAK validation + final_patch.diff
+├── runs/           # GEAK experiment tree (its own copy of the kernel)
+├── handoff.json
+└── result.json
+```
+
+They sit beside the workspace (not inside it) because GEAK copies `kernel_path`
+into its own experiment tree — nesting outputs under the workspace would recurse,
+and it keeps GEAK's scratch clear of the directory Arena scores.
+
+## Integrity boundary
+
+`geak_v4` lets GEAK edit the workspace in place, exactly like `forge` and
+`claude_code`. Integrity is enforced by the Arena harness, not the launcher:
+`main.py` snapshots the harness before the run, verifies it afterwards,
+re-materializes the perf helpers, and re-scores the kernel independently. A run
+that tampers with a protected harness path fails harness verification. This is a
+fail-closed correctness control, not an OS security sandbox; a failed run is not
+automatically rolled back. A real paid Claude workflow invocation is not part of
+the offline tests — run the one-task example with an authorized account before
+relying on this integration for a benchmark campaign.

agents/geak_v4/__init__.py

@@ -0,0 +1,4 @@
+# Copyright(C) [2026] Advanced Micro Devices, Inc. All rights reserved.
+from .launch_agent import launch_agent
+
+__all__ = ["launch_agent"]

agents/geak_v4/agent_config.yaml

@@ -0,0 +1,30 @@
+# GEAK v4 runs the deterministic kernel_workflow through Claude Code's dynamic
+# Workflow tool. Point the runner at your GEAK checkout's kernel_workflow/
+# directory; the GEAK_V4_WORKFLOW_DIR environment variable overrides this value.
+# All run artifacts are written to a hidden sibling of the Arena task workspace.
+workflow_dir: /opt/geak/kernel_workflow
+
+# Defaults from GEAK v4's supported standalone kernel workflow.
+model: claude-opus-4-8
+effort: ultracode
+budget: 6
+min_improve: 0.02
+deep_cost: 2
+use_expert_skills: false
+
+# Hard wall-clock bound for the complete multi-agent workflow.
+timeout_seconds: 43200
+
+# Newer Claude Code builds may finish detached validation work after the
+# background task notification. Keep the SDK client alive for this bounded
+# on-disk completion grace period.
+done_grace_seconds: 1800
+done_poll_seconds: 5
+
+# V1 deliberately targets existing, isolated kernels with one declared source
+# file. Author/translation tasks need a separate frozen-baseline adapter, while
+# repository/image tasks need a storage/copy-cost qualification first.
+supported_task_types:
+  - hip2hip
+  - triton2triton
+  - flydsl2flydsl