Merge pull request #970 from langchain-ai/yogesh/support-agents.md

feat: inject AGENTS.md into system prompt

apps/agent/agent/prompt.py

@@ -4,6 +4,7 @@ You are operating in a **remote Linux sandbox** at `{working_dir}`.
 
 All code execution and file operations happen in this sandbox environment.
 
+{agents_md_section}
 
 **Important:**
 - Use `{working_dir}` as your working directory for all operations
@@ -72,9 +73,20 @@ def construct_system_prompt(
     working_dir: str,
     linear_project_id: str = "",
     linear_issue_number: str = "",
+    agents_md: str = "",
 ) -> str:
+    agents_md_section = ""
+    if agents_md:
+        agents_md_section = (
+            "\nThe following text is pulled from the repository's AGENTS.md file. "
+            "It may contain specific instructions and guidelines for the agent.\n"
+            "<agents_md>\n"
+            f"{agents_md}\n"
+            "</agents_md>\n"
+        )
     return SYSTEM_PROMPT.format(
         working_dir=working_dir,
         linear_project_id=linear_project_id or "<PROJECT_ID>",
         linear_issue_number=linear_issue_number or "<ISSUE_NUMBER>",
+        agents_md_section=agents_md_section,
     )

apps/agent/agent/server.py

@@ -42,6 +42,7 @@ SANDBOX_CREATING = "__creating__"
 SANDBOX_CREATION_TIMEOUT = 180
 SANDBOX_POLL_INTERVAL = 1.0
 
+from .utils.agents_md import read_agents_md_in_sandbox
 from .utils.github import (
     git_has_uncommitted_changes,
     is_valid_git_repo,
@@ -215,7 +216,6 @@ async def get_agent(config: RunnableConfig) -> Pregel:  # noqa: PLR0915
     encrypted_token = config["configurable"].get("github_token_encrypted")
     if encrypted_token:
         github_token = decrypt_token(encrypted_token)
-        logger.debug("Decrypted GitHub token")
 
     if thread_id is None or not graph_loaded_for_execution(config):
         logger.info("No thread_id or not for execution, returning agent without sandbox")
@@ -324,9 +324,14 @@ async def get_agent(config: RunnableConfig) -> Pregel:  # noqa: PLR0915
 
     SANDBOX_BACKENDS[thread_id] = sandbox_backend
 
+    if not repo_dir:
+        msg = "Cannot proceed: no repo was cloned. Set 'repo.owner' and 'repo.name' in the configurable config"
+        raise RuntimeError(msg)
+
     linear_issue = config["configurable"].get("linear_issue", {})
     linear_project_id = linear_issue.get("linear_project_id", "")
     linear_issue_number = linear_issue.get("linear_issue_number", "")
+    agents_md = await read_agents_md_in_sandbox(sandbox_backend, repo_dir)
 
     logger.info("Returning agent with sandbox for thread %s", thread_id)
     return create_deep_agent(
@@ -335,6 +340,7 @@ async def get_agent(config: RunnableConfig) -> Pregel:  # noqa: PLR0915
             repo_dir,
             linear_project_id=linear_project_id,
             linear_issue_number=linear_issue_number,
+            agents_md=agents_md,
         ),
         tools=[http_request, fetch_url, commit_and_open_pr],
         backend=sandbox_backend,

apps/agent/agent/utils/agents_md.py

@@ -0,0 +1,34 @@
+"""Helpers for reading agent instructions from AGENTS.md."""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import shlex
+
+from deepagents.backends.protocol import SandboxBackendProtocol
+
+logger = logging.getLogger(__name__)
+
+
+async def read_agents_md_in_sandbox(
+    sandbox_backend: SandboxBackendProtocol,
+    repo_dir: str | None,
+) -> str | None:
+    """Read AGENTS.md from the repo root if it exists."""
+    if not repo_dir:
+        return None
+
+    safe_agents_path = shlex.quote(f"{repo_dir}/AGENTS.md")
+    loop = asyncio.get_event_loop()
+    result = await loop.run_in_executor(
+        None,
+        sandbox_backend.execute,
+        f"test -f {safe_agents_path} && cat {safe_agents_path}",
+    )
+    if result.exit_code != 0:
+        logger.debug("AGENTS.md not found at %s", safe_agents_path)
+        return None
+    content = result.output or ""
+    content = content.strip()
+    return content or None