diff --git a/docs/plans/2026-02-25-devcontainer-design.md b/docs/plans/2026-02-25-devcontainer-design.md
new file mode 100644
index 000000000..2c2a5e076
--- /dev/null
+++ b/docs/plans/2026-02-25-devcontainer-design.md
@@ -0,0 +1,232 @@
+# DevPod Remote Development Environment — Design
+
+**Date**: 2026-02-25
+**Status**: Approved
+
+## Goal
+
+A one-click remote development environment on Kapsule with full Claude Code
+toolchain, persistent storage, and autoscale-to-zero cost control. Accessible
+from anywhere via an "Open in DevPod" link on the GitLab project page.
+
+## Decisions
+
+| Decision | Choice | Rationale |
+|----------|--------|-----------|
+| Tool | DevPod (Loft Labs) | Zero server infra, native devcontainer.json, built-in idle timeout, provider-agnostic |
+| Dev node | GP1-L (16 vCPU, 64GB) | Fast cargo builds on 37-crate workspace, autoscales to zero |
+| GPU for dev | None (CPU-only) | cargo check/clippy/test are CPU-bound; GPU training uses H100 via CI/jobs |
+| GPU for training | H100 (existing pool) | Stays autoscaled to zero, wakes for CI and ad-hoc training jobs |
+| Connection | Terminal + Claude Code | No VS Code/JetBrains server overhead |
+| Persistence | Block storage PVCs | Survive pod restarts, node scale-down, Claude Code auto-updates |
+| GitLab integration | Project badge + CI image build | "Open in DevPod" button, image rebuilt on .devcontainer/ changes |
+| Eliminated | GitLab Workspaces (no GPU, Devfile only), Coder (overkill), raw SSH (manual lifecycle) |
+
+## Infrastructure
+
+### New Kapsule Pool: `gpu-dev`
+
+```hcl
+resource "scaleway_k8s_pool" "gpu_dev" {
+  name        = "gpu-dev"
+  node_type   = "GP1-L"     # 16 vCPU, 64GB RAM
+  min_size    = 0
+  max_size    = 1
+  autoscaling = true
+  autohealing = true
+
+  lifecycle {
+    ignore_changes = [size]
+  }
+}
+```
+
+No GPU taint, no special tolerations. The node selector
+`k8s.scaleway.com/pool-name=gpu-dev` directs the dev pod here.
+
+### Persistent Volumes
+
+| PVC | Size | StorageClass | Mount | Purpose |
+|-----|------|-------------|-------|---------|
+| `dev-home-pvc` | 50Gi | scw-bssd | `/home/dev` | Claude Code + ~/.claude/ + cargo registry + sccache cache + shell/git config |
+| `training-data-pvc` | 10Gi (existing) | scw-bssd | `/data/training` | Databento .dbn.zst futures data (read-only) |
+
+DevPod also creates a workspace volume for `/workspace` (the repo checkout).
+
+The 50Gi home volume is intentionally large: cargo registry + sccache for 37
+crates, Claude Code plugins/MCP caches, and future datasets or model checkpoints.
+The PVC can be resized, mounted in other pods, or snapshotted as needed.
+
+## Devcontainer Image
+
+Built by CI, pushed to `git.fxhnt.ai:5050/root/foxhunt/devcontainer:latest`.
+
+### Base
+
+Extends `Dockerfile.ci-builder` (CUDA 12.4, Rust 1.89, protoc, sccache, clippy).
+
+### Additional Layers
+
+**Dev tooling**:
+- Node.js 22 LTS (Claude Code runtime)
+- Claude Code CLI (`npm install -g @anthropic-ai/claude-code`)
+- Python 3.12 + pip (scripts, data tooling)
+- kubectl, helm (cluster management)
+- scw CLI (Scaleway management)
+- glab (GitLab CLI)
+- terraform, terragrunt (infra management)
+- jq, yq, ripgrep, fd-find (CLI essentials)
+- zsh + oh-my-zsh (shell)
+- git-lfs
+
+**Environment**:
+- Non-root user `dev` with passwordless sudo
+- Cargo env in `/home/dev/.cargo`
+- sccache configured (local disk cache on PVC)
+- `SQLX_OFFLINE=true`
+- `CUDA_COMPUTE_CAP=90`
+
+## `.devcontainer/` Structure
+
+```
+.devcontainer/
+├── devcontainer.json      # DevPod + devcontainer spec
+├── Dockerfile             # Extends ci-builder with dev tooling
+└── pod-template.yaml      # K8s pod overrides (PVC mounts, node selector)
+```
+
+### devcontainer.json
+
+```jsonc
+{
+  "name": "foxhunt",
+  "build": {
+    "dockerfile": "Dockerfile",
+    "context": ".."
+  },
+  "remoteUser": "dev",
+  "containerEnv": {
+    "SQLX_OFFLINE": "true",
+    "CUDA_COMPUTE_CAP": "90"
+  },
+  "postStartCommand": "claude --version && cargo --version && kubectl cluster-info 2>/dev/null || true",
+  "customizations": {
+    "devpod": {
+      "provider": "kubernetes"
+    }
+  }
+}
+```
+
+PVC mounts and node selector are handled by `pod-template.yaml` (DevPod's
+`POD_MANIFEST_TEMPLATE` option), since devcontainer.json cannot express k8s-
+specific volume claims.
+
+## DevPod Provider Configuration
+
+One-time setup on the developer's machine (`scripts/devpod-setup.sh`):
+
+```bash
+devpod provider add kubernetes
+devpod provider set kubernetes \
+  KUBERNETES_NAMESPACE=foxhunt \
+  NODE_SELECTOR="k8s.scaleway.com/pool-name=gpu-dev" \
+  DISK_SIZE=10Gi \
+  STORAGE_CLASS=scw-bssd \
+  INACTIVITY_TIMEOUT=30m \
+  IMAGE_PULL_SECRETS=gitlab-registry \
+  POD_MANIFEST_TEMPLATE=.devcontainer/pod-template.yaml
+```
+
+## GitLab Integration
+
+### "Open in DevPod" Badge
+
+Added as a GitLab project badge (Settings → General → Badges):
+
+```
+Name:  Open in DevPod
+Link:  https://devpod.sh/open#ssh://git@100.90.76.85:2222/root/foxhunt.git
+Image: https://devpod.sh/assets/open-in-devpod.svg
+```
+
+Also usable as:
+- CLI: `devpod up git@100.90.76.85:2222/root/foxhunt.git`
+- Bookmark in any browser
+- Shared link in docs/chat
+
+### CI Image Build
+
+New job in `.gitlab-ci.yml`:
+
+```yaml
+build-devcontainer:
+  stage: build
+  rules:
+    - changes: [".devcontainer/**"]
+  script:
+    # Kaniko build + push to GitLab registry
+```
+
+Only triggers when `.devcontainer/` files change. Uses existing Kaniko
+pattern from the service image builds.
+
+## Session Lifecycle
+
+```
+Click "Open in DevPod" (or `devpod up`)
+  → DevPod creates pod targeting gpu-dev pool
+  → Kapsule autoscaler provisions GP1-L node (~2-3 min)
+  → Pod starts, mounts PVCs (home + training data)
+  → SSH tunnel established
+  → Terminal opens
+
+Active development
+  → cargo check, claude, kubectl, glab, terraform
+  → All writes go to persistent /home/dev PVC
+
+Idle 30 minutes (no SSH activity)
+  → DevPod stops pod
+  → Node idle 10 min
+  → Kapsule autoscaler removes GP1-L node
+  → Cost: €0
+
+Return (click link or `devpod up` again)
+  → Node provisions (~2-3 min)
+  → Pod starts, mounts SAME PVCs
+  → Claude Code config, cargo cache, shell history intact
+  → Incremental cargo builds (sccache hits)
+```
+
+## GPU Access (On-Demand)
+
+The dev pod is CPU-only. GPU access for training/inference:
+
+- **CI training jobs** → H100 pool (existing, autoscaled to zero)
+- **Ad-hoc training** → `kubectl apply -f infra/k8s/training/job-template.yaml`
+  from inside the dev pod
+- **Future**: second DevPod provider profile targeting L4 or H100 pool with
+  `nvidia.com/gpu=1` for interactive GPU sessions
+
+## Cost Estimate
+
+| Scenario | Node | Rate | Daily (8hr) | Monthly |
+|----------|------|------|-------------|---------|
+| Active dev | GP1-L | ~€0.24/hr | ~€2 | ~€44 |
+| Idle | — | €0 | €0 | €0 |
+| CI build (H100) | H100-1-80G | ~€3-4/hr | per-job | per-job |
+
+No conflict between dev and CI — separate node pools.
+
+## Files to Create/Modify
+
+| File | Action |
+|------|--------|
+| `.devcontainer/devcontainer.json` | Create |
+| `.devcontainer/Dockerfile` | Create |
+| `.devcontainer/pod-template.yaml` | Create |
+| `infra/modules/kapsule/main.tf` | Add gpu-dev pool resource |
+| `infra/modules/kapsule/variables.tf` | Add dev pool variables |
+| `infra/live/production/kapsule/terragrunt.hcl` | Enable dev pool, set type |
+| `.gitlab-ci.yml` | Add build-devcontainer job |
+| `scripts/devpod-setup.sh` | One-time provider config script |