Add agents-md plugin: auto-load AGENTS.md at session start (#6235)

plugins/README.md

@@ -12,6 +12,7 @@ Learn more in the [official plugins documentation](https://docs.claude.com/en/do
 
 | Name | Description | Contents |
 |------|-------------|----------|
+| [agents-md](./agents-md/) | Loads `AGENTS.md` files into context at session start, making Claude Code compatible with the cross-tool agent instruction standard used by Cursor, Codex, and Amp ([#6235](https://github.com/anthropics/claude-code/issues/6235)) | **Hook:** SessionStart - Injects `AGENTS.md` content as `additionalContext`, mirroring `CLAUDE.md` discovery |
 | [agent-sdk-dev](./agent-sdk-dev/) | Development kit for working with the Claude Agent SDK | **Command:** `/new-sdk-app` - Interactive setup for new Agent SDK projects<br>**Agents:** `agent-sdk-verifier-py`, `agent-sdk-verifier-ts` - Validate SDK applications against best practices |
 | [claude-opus-4-5-migration](./claude-opus-4-5-migration/) | Migrate code and prompts from Sonnet 4.x and Opus 4.1 to Opus 4.5 | **Skill:** `claude-opus-4-5-migration` - Automated migration of model strings, beta headers, and prompt adjustments |
 | [code-review](./code-review/) | Automated PR code review using multiple specialized agents with confidence-based scoring to filter false positives | **Command:** `/code-review` - Automated PR review workflow<br>**Agents:** 5 parallel Sonnet agents for CLAUDE.md compliance, bug detection, historical context, PR history, and code comments |
@@ -25,6 +26,7 @@ Learn more in the [official plugins documentation](https://docs.claude.com/en/do
 | [pr-review-toolkit](./pr-review-toolkit/) | Comprehensive PR review agents specializing in comments, tests, error handling, type design, code quality, and code simplification | **Command:** `/pr-review-toolkit:review-pr` - Run with optional review aspects (comments, tests, errors, types, code, simplify, all)<br>**Agents:** `comment-analyzer`, `pr-test-analyzer`, `silent-failure-hunter`, `type-design-analyzer`, `code-reviewer`, `code-simplifier` |
 | [ralph-wiggum](./ralph-wiggum/) | Interactive self-referential AI loops for iterative development. Claude works on the same task repeatedly until completion | **Commands:** `/ralph-loop`, `/cancel-ralph` - Start/stop autonomous iteration loops<br>**Hook:** Stop - Intercepts exit attempts to continue iteration |
 | [security-guidance](./security-guidance/) | Security reminder hook that warns about potential security issues when editing files | **Hook:** PreToolUse - Monitors 9 security patterns including command injection, XSS, eval usage, dangerous HTML, pickle deserialization, and os.system calls |
+| [tmp-cwd-cleanup](./tmp-cwd-cleanup/) | Cleans up orphaned `/tmp/claude-*-cwd` files on session exit, working around a Bash tool memory leak ([#8856](https://github.com/anthropics/claude-code/issues/8856)) | **Hook:** Stop - Deletes accumulated working-directory tracking files owned by the current user |
 
 ## Installation
 

plugins/agents-md/.claude-plugin/plugin.json

@@ -0,0 +1,9 @@
+{
+  "name": "agents-md",
+  "version": "1.0.0",
+  "description": "Automatically loads AGENTS.md files into context at session start, making Claude Code compatible with cross-tool agent instruction files used by Cursor, Codex, Amp, and others",
+  "author": {
+    "name": "Claude Code Community",
+    "email": ""
+  }
+}

plugins/agents-md/README.md

@@ -0,0 +1,64 @@
+# agents-md
+
+Automatically loads `AGENTS.md` files into Claude Code's context at session start, making it compatible with the cross-tool agent instruction standard used by Cursor, OpenAI Codex, Amp, and others.
+
+**Fixes:** [#6235 – Feature Request: Support AGENTS.md](https://github.com/anthropics/claude-code/issues/6235)
+
+## The problem
+
+Claude Code reads `CLAUDE.md` (and `CLAUDE.local.md`) from the project directory tree to understand project-specific instructions. Most other AI coding tools use `AGENTS.md` for the same purpose — this is coalescing into a [cross-tool standard](https://agents.md/).
+
+Without this plugin, teams that use both Claude Code and other tools must either:
+- Maintain two separate instruction files (`CLAUDE.md` and `AGENTS.md`) in sync, or
+- Symlink one to the other (fragile on Windows, confusing for contributors)
+
+## What this plugin does
+
+Registers a `SessionStart` hook that mirrors Claude Code's own `CLAUDE.md` discovery logic:
+
+1. Walks up the directory tree from the project root, collecting every `AGENTS.md` found.
+2. Also checks `~/.claude/AGENTS.md` for user-level instructions.
+3. Injects all found content as `additionalContext` — the same mechanism used internally for `CLAUDE.md`.
+
+The file lookup order (lowest → highest priority) matches `CLAUDE.md` precedence:
+
+```
+~/.claude/AGENTS.md          ← user-level (lowest priority)
+/path/to/repo/AGENTS.md      ← repo root
+/path/to/repo/subdir/AGENTS.md  ← subdirectory (highest priority)
+```
+
+## Installation
+
+Install via the `/plugin` command inside Claude Code:
+
+```
+/plugin install agents-md
+```
+
+Or add it manually to your `.claude/settings.json`:
+
+```json
+{
+  "plugins": ["agents-md"]
+}
+```
+
+## Usage
+
+Just drop an `AGENTS.md` in your project root (the same place you'd put `CLAUDE.md`) and start a Claude Code session. The file is loaded automatically — no configuration needed.
+
+```
+my-project/
+├── AGENTS.md      ← loaded by this plugin (and by Cursor, Codex, Amp, …)
+├── CLAUDE.md      ← loaded natively by Claude Code
+└── src/
+```
+
+Both files are loaded; `CLAUDE.md` instructions take precedence where they overlap.
+
+## Hook
+
+| Event | Effect |
+|-------|--------|
+| `SessionStart` | Finds and injects `AGENTS.md` file(s) as `additionalContext` |

plugins/agents-md/hooks/hooks.json

@@ -0,0 +1,15 @@
+{
+  "description": "Loads AGENTS.md files into context at session start, mirroring how Claude Code handles CLAUDE.md",
+  "hooks": {
+    "SessionStart": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "python3 ${CLAUDE_PLUGIN_ROOT}/hooks/load_agents_md.py"
+          }
+        ]
+      }
+    ]
+  }
+}

plugins/agents-md/hooks/load_agents_md.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+"""SessionStart hook: inject AGENTS.md files as additionalContext.
+
+Claude Code natively reads CLAUDE.md (and CLAUDE.local.md) from the project
+tree and the user's ~/.claude/ directory. Many other AI coding tools — Cursor,
+OpenAI Codex, Amp, and others — use AGENTS.md for the same purpose, creating
+a cross-tool standard.
+
+This hook mirrors Claude Code's own CLAUDE.md discovery logic:
+  1. Walk up from the project root, collecting every AGENTS.md found.
+  2. Also check ~/.claude/AGENTS.md for user-level instructions.
+
+All found files are concatenated and injected via additionalContext so Claude
+sees them exactly as if they had been written in CLAUDE.md.
+
+See: https://github.com/anthropics/claude-code/issues/6235
+     https://agents.md/
+"""
+
+import json
+import os
+import sys
+from pathlib import Path
+
+
+FILENAME = "AGENTS.md"
+# Limit how many directory levels we climb (same spirit as CLAUDE.md behaviour).
+MAX_DEPTH = 20
+
+
+def find_agents_md_files(start_dir: Path) -> list[Path]:
+    """Walk upward from start_dir collecting AGENTS.md files (innermost first)."""
+    found: list[Path] = []
+    current = start_dir.resolve()
+
+    for _ in range(MAX_DEPTH):
+        candidate = current / FILENAME
+        if candidate.is_file():
+            found.append(candidate)
+        parent = current.parent
+        if parent == current:
+            break  # reached filesystem root
+        current = parent
+
+    return found
+
+
+def main() -> None:
+    # Consume stdin (required by hook protocol).
+    try:
+        hook_input = json.load(sys.stdin)
+    except Exception:
+        hook_input = {}
+
+    # Determine the project root. The hook runs with cwd set to the project
+    # directory; fall back to CLAUDE_PROJECT_ROOT if provided.
+    project_root_env = os.environ.get("CLAUDE_PROJECT_ROOT", "")
+    start_dir = Path(project_root_env) if project_root_env else Path.cwd()
+
+    # 1. Collect project-tree AGENTS.md files (closest → furthest from root).
+    project_files = find_agents_md_files(start_dir)
+
+    # 2. User-level AGENTS.md at ~/.claude/AGENTS.md (lowest priority).
+    user_file = Path.home() / ".claude" / FILENAME
+    user_files = [user_file] if user_file.is_file() else []
+
+    # Final order: user-level first (lowest priority), then project files
+    # from outermost → innermost (matching how CLAUDE.md precedence works).
+    all_files = user_files + list(reversed(project_files))
+
+    if not all_files:
+        # Nothing to inject — exit silently.
+        sys.exit(0)
+
+    # Read and concatenate.
+    sections: list[str] = []
+    for path in all_files:
+        try:
+            content = path.read_text(encoding="utf-8").strip()
+            if content:
+                sections.append(
+                    f"<!-- AGENTS.md: {path} -->\n{content}"
+                )
+        except OSError:
+            pass  # Unreadable file — skip silently.
+
+    if not sections:
+        sys.exit(0)
+
+    combined = "\n\n".join(sections)
+    context = (
+        "The following instructions come from AGENTS.md file(s) found in this "
+        "project. AGENTS.md is a cross-tool standard (used by Cursor, Codex, "
+        "Amp, and others) equivalent to CLAUDE.md. Treat these instructions "
+        "with the same weight as CLAUDE.md:\n\n" + combined
+    )
+
+    output = {
+        "hookSpecificOutput": {
+            "hookEventName": "SessionStart",
+            "additionalContext": context,
+        }
+    }
+    print(json.dumps(output))
+    sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()

plugins/tmp-cwd-cleanup/.claude-plugin/plugin.json

@@ -0,0 +1,9 @@
+{
+  "name": "tmp-cwd-cleanup",
+  "version": "1.0.0",
+  "description": "Cleans up orphaned /tmp/claude-*-cwd working directory tracking files on session stop, working around a memory leak in the Bash tool (issue #8856)",
+  "author": {
+    "name": "Claude Code Community",
+    "email": ""
+  }
+}

plugins/tmp-cwd-cleanup/README.md

@@ -0,0 +1,41 @@
+# tmp-cwd-cleanup
+
+A workaround plugin for [issue #8856](https://github.com/anthropics/claude-code/issues/8856): the Bash tool creates a `/tmp/claude-<hex>-cwd` file after every shell command to track the working directory, but never deletes it. On active systems these files accumulate into the thousands and eventually slow down `/tmp` lookups.
+
+## What it does
+
+Registers a **Stop hook** that deletes all `/tmp/claude-*-cwd` files owned by the current user when the Claude Code session ends.
+
+```
+/tmp/claude-02a6-cwd   (22 bytes — deleted on exit)
+/tmp/claude-1f3b-cwd   (22 bytes — deleted on exit)
+...
+```
+
+## Installation
+
+Install via the `/plugin` command inside Claude Code:
+
+```
+/plugin install tmp-cwd-cleanup
+```
+
+Or add it manually to your `.claude/settings.json`:
+
+```json
+{
+  "plugins": ["tmp-cwd-cleanup"]
+}
+```
+
+## Notes
+
+- Only removes files owned by the **current user** (safe on shared systems).
+- Exits `0` so it never blocks session teardown.
+- This plugin is a stopgap; the proper fix is to call `unlinkSync` in the Bash tool implementation immediately after reading the cwd file ([upstream tracking issue #8856](https://github.com/anthropics/claude-code/issues/8856)).
+
+## Hook
+
+| Event | Matcher | Effect |
+|-------|---------|--------|
+| `Stop` | _(all)_ | Deletes `/tmp/claude-*-cwd` files owned by current user |

plugins/tmp-cwd-cleanup/hooks/cleanup_cwd_files.py

@@ -0,0 +1,53 @@
+#!/usr/bin/env python3
+"""Stop hook: clean up orphaned /tmp/claude-*-cwd files.
+
+The Bash tool writes a small temp file to /tmp/claude-<hex>-cwd after each
+command to capture the resulting working directory, but never deletes it.
+On active systems these files accumulate into the thousands and eventually
+slow down /tmp lookups (see issue #8856).
+
+This hook runs on session Stop and removes any such files owned by the
+current user, providing a clean-up path until the upstream fix lands.
+"""
+
+import glob
+import json
+import os
+import sys
+
+
+def main() -> None:
+    # Consume stdin (required by the hook protocol) but we don't need it.
+    try:
+        json.load(sys.stdin)
+    except Exception:
+        pass
+
+    uid = os.getuid() if hasattr(os, "getuid") else None  # no-op on Windows
+    removed = 0
+    errors = 0
+
+    for path in glob.glob("/tmp/claude-*-cwd"):
+        try:
+            # Only remove files owned by the current user (safety check).
+            if uid is not None and os.stat(path).st_uid != uid:
+                continue
+            if os.path.isfile(path):
+                os.unlink(path)
+                removed += 1
+        except OSError:
+            errors += 1
+
+    # Exit 0 so the Stop event is never blocked.
+    # Optionally surface a brief summary via systemMessage.
+    if removed > 0:
+        result = {"systemMessage": f"tmp-cwd-cleanup: removed {removed} orphaned file(s)."}
+        if errors:
+            result["systemMessage"] += f" ({errors} error(s) skipped)"
+        print(json.dumps(result))
+
+    sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()

plugins/tmp-cwd-cleanup/hooks/hooks.json

@@ -0,0 +1,15 @@
+{
+  "description": "Cleans up orphaned /tmp/claude-*-cwd working directory tracking files when the session stops",
+  "hooks": {
+    "Stop": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "python3 ${CLAUDE_PLUGIN_ROOT}/hooks/cleanup_cwd_files.py"
+          }
+        ]
+      }
+    ]
+  }
+}