@tangle-network/sandbox-cli 0.9.4-develop.20260629071229.145161 → 0.9.4-develop.20260701061438.a81baba

package/README.md

@@ -119,6 +119,15 @@ tangle agent task sbx_123 "triage issues" \
   --connection github:github.issues.search,github.issues.create
 tangle agent task sbx_123 "file the report" --connection github:* --allow-writes
 
+# temporary GPU for an eval; omitted provider picks the cheapest configured cloud
+tangle sandbox gpu run sbx_123 \
+  --accelerator-kind nvidia-3090 \
+  --accelerator-memory 24000 \
+  --max-spend-usd 1 \
+  --max-lifetime 900 \
+  --idle-timeout 120 \
+  -- python eval.py
+
 # state and operations
 tangle secret list
 tangle snapshot list sbx_123
@@ -133,7 +142,7 @@ grant modes, `--allow-writes`, and `permissions revert-writes`.
 
 ## Provisioning Coverage
 
-`tangle sandbox create` exposes the full SDK provisioning surface, including image/environment, resources (CPU, memory, disk), lifetime and idle timeout, driver and backend selection (`opencode`, `claude-code`, `codex`, `cursor`, `amp`, and other registry backends), SSH and web terminal, env and secret injection, metadata, initial permissions, git clone, tool pre-install, BYOS3 storage, snapshot restore, and outbound network controls. Run `tangle sandbox create --help` for the canonical flag list.
+`tangle sandbox create` exposes the full SDK provisioning surface, including image/environment, resources (CPU, memory, disk), temporary GPU leases, lifetime and idle timeout, driver and backend selection (`opencode`, `claude-code`, `codex`, `cursor`, `amp`, and other registry backends), SSH and web terminal, env and secret injection, metadata, initial permissions, git clone, tool pre-install, BYOS3 storage, snapshot restore, and outbound network controls. Run `tangle sandbox create --help` for the canonical flag list.
 
 ## Limitations
 

package/SKILL.md

@@ -67,6 +67,44 @@ tangle sandbox expose <id> --port 3000
 tangle sandbox urls <id>
 ```
 
+## Temporary GPUs
+
+Use a normal sandbox first, then attach a GPU only for the work that needs it.
+Omit `--provider` to use the cheapest configured GPU cloud.
+Always set a spend cap and lifetime.
+
+```bash
+# One-shot eval: attach GPU, run command, always destroy the lease.
+tangle sandbox gpu run <sandbox-id> \
+  --accelerator-kind nvidia-3090 \
+  --accelerator-memory 24000 \
+  --max-spend-usd 1 \
+  --max-lifetime 900 \
+  --idle-timeout 120 \
+  -- python eval.py
+
+# Manual lifecycle when the agent needs multiple GPU commands.
+tangle sandbox gpu attach <sandbox-id> \
+  --accelerator-kind nvidia-3090 \
+  --accelerator-memory 24000 \
+  --max-spend-usd 1 \
+  --max-lifetime 900 \
+  --idle-timeout 120
+tangle sandbox gpu exec <sandbox-id> <gpu-lease-id> -- python eval.py
+tangle sandbox gpu detach <sandbox-id> <gpu-lease-id>
+```
+
+For fleets, put the same GPU lease cap on every worker:
+
+```bash
+tangle fleet create --count 4 \
+  --accelerator-kind nvidia-3090 \
+  --accelerator-memory 24000 \
+  --gpu-max-spend-usd 1 \
+  --gpu-max-lifetime 900 \
+  --gpu-idle-timeout 120
+```
+
 ## Hub — Full Workflow
 
 Hub lets agents use connected provider tools (GitHub, etc.) through Tangle without seeing provider OAuth tokens.