Tb-2

AGENTS.md

@@ -91,7 +91,8 @@ your response.
   read-only dependency unless coordinating with maintainers.
 - `examples/` — Runnable training recipes grouped by task type
   (`math/`, `code/`, `math-multi-agent/`, `code-multi-agent/`,
-  `alfworld/`, `webshop/`, `search/`, `math-efficient-data/`), plus
+  `terminal-bench/`, `alfworld/`, `webshop/`, `search/`,
+  `math-efficient-data/`), plus
   shared helpers in `_common/` and `launch_trainer.py`. Each recipe
   ships a `yaml/` directory of configs and a `scripts/` directory of
   numbered launch scripts.

README.md

@@ -64,6 +64,7 @@ AstraFlow currently supports the following recipes. Check the [documentation](ht
 | [`math-efficient-data/`](examples/math-efficient-data/) | Composable data algorithms — GRESO, dynamic sampling, buffer replay |
 | [`code/`](examples/code/) | Code-generation RL — Qwen3-8B, M2PO |
 | [`code-multi-agent/`](examples/code-multi-agent/) | Codegen + verifier competitive coding |
+| [`terminal-bench/`](examples/terminal-bench/) | Harbor-backed Terminal-Bench eval and RL recipes |
 | [`search/`](examples/search/) | Search-augmented agent training with local retrieval |
 | [`alfworld/`](examples/alfworld/) | ALFWorld embodied household agent |
 | [`webshop/`](examples/webshop/) | WebShop web-navigation shopping agent |

astraflow/core/workflow/__init__.py

@@ -21,6 +21,7 @@
 import astraflow.core.workflow.impl.solve_and_check
 import astraflow.core.workflow.impl.sep_solve_and_check
 import astraflow.core.workflow.impl.solve_and_verify
+import astraflow.core.workflow.impl.terminal_bench_harbor
 import astraflow.core.workflow.impl.actor_and_verify
 import astraflow.core.workflow.impl.rlvr
 import astraflow.core.workflow.impl.sm_lg_router