Add AGS rollout buffer generator for SWE coding-agent RL

.github/workflows/pr-test.yml

@@ -744,6 +744,18 @@ jobs:
             {
               "num_gpus": 0,
               "test_file": "test_agent/test_agent_rollout_cpu.py"
+            },
+            {
+              "num_gpus": 0,
+              "test_file": "test_rollout_buffer/test_ags_prompt_source.py"
+            },
+            {
+              "num_gpus": 0,
+              "test_file": "test_rollout_buffer/test_ags_empty_patch_guard.py"
+            },
+            {
+              "num_gpus": 0,
+              "test_file": "test_rollout_buffer/test_ags_generator.py"
             }
           ]
     defaults:
@@ -774,6 +786,8 @@ jobs:
           pip install torch --index-url https://download.pytorch.org/whl/cpu
           pip install pytest numpy packaging pyyaml omegaconf tqdm httpx requests ray pybase64 pylatexenc sympy aiohttp pillow safetensors
 
+          pip install openai transformers wandb fastapi uvicorn
+
 
       - name: Install
         shell: bash

.github/workflows/pr-test.yml.j2

@@ -97,11 +97,15 @@
       'label': 'run-ci-agent',
       'always': True,
       'cpu': True,
+      'extra_pip_deps': 'openai transformers wandb fastapi uvicorn',
       'tests': [
         {'test_file': 'test_agent/test_trajectory_manager_branching.py', 'num_gpus': 0},
         {'test_file': 'test_agent/test_adapters.py', 'num_gpus': 0},
         {'test_file': 'test_agent/test_harness.py', 'num_gpus': 0},
         {'test_file': 'test_agent/test_agent_rollout_cpu.py', 'num_gpus': 0},
+        {'test_file': 'test_rollout_buffer/test_ags_prompt_source.py', 'num_gpus': 0},
+        {'test_file': 'test_rollout_buffer/test_ags_empty_patch_guard.py', 'num_gpus': 0},
+        {'test_file': 'test_rollout_buffer/test_ags_generator.py', 'num_gpus': 0},
       ],
     },
 

examples/coding_agent_rl_ags/eval_cbc_qwen35_35b_a3b_swe_2nodes.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+# Score one checkpoint on SWE-bench Verified with the CodeBuddy Code (cbc)
+# harness, 2 nodes, no training.
+#
+# Thin wrapper: the launcher logic lives in eval_cc_qwen35_35b_a3b_swe_2nodes.sh
+# and is harness-agnostic -- SWE_AGENT selects the harness, and that script sets
+# both harnesses' EXTRA_ARGS/EXTRA_ENVS knobs -- so there is one copy to keep correct.
+#
+# Every variable the CC eval script reads works here too:
+#   LOAD_DIR=<ckpt_root> CKPT_STEP=99 EVAL_DATA=<jsonl> \
+#     bash examples/coding_agent_rl_ags/eval_cbc_qwen35_35b_a3b_swe_2nodes.sh
+#
+# Omit LOAD_DIR/CKPT_STEP to score the untrained HF model as the baseline.
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
+
+export SWE_AGENT=codebuddy_code
+export EXP_TAG="${EXP_TAG:-coding_agent_rl_ags_eval_cbc_qwen35_35b_a3b${CKPT_STEP:+_step${CKPT_STEP}}}"
+
+exec bash "${SCRIPT_DIR}/eval_cc_qwen35_35b_a3b_swe_2nodes.sh" "$@"

examples/coding_agent_rl_ags/run_cbc_qwen35_35b_a3b_swe_2nodes.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+# GRPO training on SWE-rebench with the CodeBuddy Code (cbc) harness, 2 nodes.
+#
+# Thin wrapper: the launcher logic lives in run_cc_qwen35_35b_a3b_swe_2nodes.sh
+# and is harness-agnostic -- SWE_AGENT selects the harness, and that script sets
+# both harnesses' EXTRA_ARGS/EXTRA_ENVS knobs. Duplicating ~370 lines per harness
+# would guarantee the two copies drift, which is the failure this rename cleans up.
+#
+# Everything the CC script accepts works here too, e.g.
+#   PROMPT_DATA=... NUM_ROLLOUT=50 bash run_cbc_qwen35_35b_a3b_swe_2nodes.sh
+#
+# Usage (from a long-lived shell / tmux on the head node, inside the container):
+#   bash examples/coding_agent_rl_ags/run_cbc_qwen35_35b_a3b_swe_2nodes.sh
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
+
+export SWE_AGENT=codebuddy_code
+export EXP_TAG="${EXP_TAG:-coding_agent_rl_ags_cbc_qwen35_35b_a3b_2nodes}"
+
+exec bash "${SCRIPT_DIR}/run_cc_qwen35_35b_a3b_swe_2nodes.sh" "$@"

examples/coding_agent_rl_ags/run_cbc_qwen35_35b_a3b_swe_4nodes.sh

@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+# GRPO training on SWE-rebench with the CodeBuddy Code (cbc) harness, 4 nodes.
+#
+# Thin wrapper around the harness-agnostic 4-node launcher; see
+# run_cbc_qwen35_35b_a3b_swe_2nodes.sh for why these are wrappers rather than
+# copies.
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
+
+export SWE_AGENT=codebuddy_code
+export EXP_TAG="${EXP_TAG:-coding_agent_rl_ags_cbc_qwen35_35b_a3b_4nodes}"
+
+exec bash "${SCRIPT_DIR}/run_cc_qwen35_35b_a3b_swe_4nodes.sh" "$@"