fix: prevent settings.json race that drops MLflow env vars

.gitignore

@@ -27,3 +27,6 @@ uv.lock
 
 # Codex skills are generated at runtime by setup_codex.py from .claude/skills/
 .agents/
+
+# Agent-plane reference clone
+agent-plane-ref/

app.py

@@ -293,20 +293,24 @@ def _configure_all_cli_auth(token):
     else:
         anthropic_base_url = f"{databricks_host}/serving-endpoints/anthropic"
 
-    settings = {
-        "env": {
-            "ANTHROPIC_MODEL": os.environ.get("ANTHROPIC_MODEL", "databricks-claude-opus-4-7"),
-            "ANTHROPIC_BASE_URL": anthropic_base_url,
-            "ANTHROPIC_AUTH_TOKEN": token,
-            "ANTHROPIC_DEFAULT_OPUS_MODEL": "databricks-claude-opus-4-7",
-            "ANTHROPIC_DEFAULT_SONNET_MODEL": "databricks-claude-sonnet-4-6",
-            "ANTHROPIC_DEFAULT_HAIKU_MODEL": "databricks-claude-haiku-4-5",
-            "ANTHROPIC_CUSTOM_HEADERS": "x-databricks-use-coding-agent-mode: true",
-            "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
-        }
-    }
-
+    # Read-merge-write to preserve env vars from other setup scripts (e.g. setup_mlflow.py)
     settings_path = os.path.join(claude_dir, "settings.json")
+    try:
+        with open(settings_path) as f:
+            settings = json.load(f)
+    except (FileNotFoundError, json.JSONDecodeError):
+        settings = {}
+
+    settings.setdefault("env", {})
+    settings["env"]["ANTHROPIC_MODEL"] = os.environ.get("ANTHROPIC_MODEL", "databricks-claude-opus-4-7")
+    settings["env"]["ANTHROPIC_BASE_URL"] = anthropic_base_url
+    settings["env"]["ANTHROPIC_AUTH_TOKEN"] = token
+    settings["env"]["ANTHROPIC_DEFAULT_OPUS_MODEL"] = "databricks-claude-opus-4-7"
+    settings["env"]["ANTHROPIC_DEFAULT_SONNET_MODEL"] = "databricks-claude-sonnet-4-6"
+    settings["env"]["ANTHROPIC_DEFAULT_HAIKU_MODEL"] = "databricks-claude-haiku-4-5"
+    settings["env"]["ANTHROPIC_CUSTOM_HEADERS"] = "x-databricks-use-coding-agent-mode: true"
+    settings["env"]["CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS"] = "1"
+
     with open(settings_path, "w") as f:
         json.dump(settings, f, indent=2)
 
@@ -373,7 +377,6 @@ def run_setup():
         ("gemini",     ["uv", "run", "python", "setup_gemini.py"]),
         ("hermes",     ["uv", "run", "python", "setup_hermes.py"]),
         ("databricks", ["uv", "run", "python", "setup_databricks.py"]),
-        ("mlflow",     ["uv", "run", "python", "setup_mlflow.py"]),
     ]
 
     with ThreadPoolExecutor(max_workers=len(parallel_steps)) as executor:
@@ -383,6 +386,11 @@ def run_setup():
         ]
         wait(futures)
 
+    # --- MLflow setup runs AFTER claude setup to avoid settings.json race ---
+    # setup_mlflow.py merges env vars into ~/.claude/settings.json which
+    # setup_claude.py also writes; running sequentially prevents clobbering.
+    _run_step("mlflow", ["uv", "run", "python", "setup_mlflow.py"])
+
     # Sync latest token into all CLI configs — covers the race where PAT
     # rotation happened while a setup script was still installing (the
     # rotation's update_cli_tokens() call silently skips missing config files).

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "coda"
-version = "0.17.4"
+version = "0.18.1"
 description = "CoDA - Coding Agents on Databricks Apps"
 requires-python = ">=3.10"
 dependencies = [

setup_claude.py

@@ -29,20 +29,27 @@
         anthropic_base_url = f"{databricks_host}/serving-endpoints/anthropic"
         print(f"Using Databricks Host: {databricks_host}")
 
-    settings = {
-        "env": {
-            "ANTHROPIC_MODEL": os.environ.get("ANTHROPIC_MODEL", "databricks-claude-opus-4-7"),
-            "ANTHROPIC_BASE_URL": anthropic_base_url,
-            "ANTHROPIC_AUTH_TOKEN": token,
-            "ANTHROPIC_DEFAULT_OPUS_MODEL": "databricks-claude-opus-4-7",
-            "ANTHROPIC_DEFAULT_SONNET_MODEL": "databricks-claude-sonnet-4-6",
-            "ANTHROPIC_DEFAULT_HAIKU_MODEL": "databricks-claude-haiku-4-5",
-            "ANTHROPIC_CUSTOM_HEADERS": "x-databricks-use-coding-agent-mode: true",
-            "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
-        }
-    }
-
     settings_path = claude_dir / "settings.json"
+
+    # Read-merge-write to preserve env vars from other setup scripts (e.g. setup_mlflow.py)
+    if settings_path.exists():
+        try:
+            settings = json.loads(settings_path.read_text())
+        except (json.JSONDecodeError, OSError):
+            settings = {}
+    else:
+        settings = {}
+
+    settings.setdefault("env", {})
+    settings["env"]["ANTHROPIC_MODEL"] = os.environ.get("ANTHROPIC_MODEL", "databricks-claude-opus-4-7")
+    settings["env"]["ANTHROPIC_BASE_URL"] = anthropic_base_url
+    settings["env"]["ANTHROPIC_AUTH_TOKEN"] = token
+    settings["env"]["ANTHROPIC_DEFAULT_OPUS_MODEL"] = "databricks-claude-opus-4-7"
+    settings["env"]["ANTHROPIC_DEFAULT_SONNET_MODEL"] = "databricks-claude-sonnet-4-6"
+    settings["env"]["ANTHROPIC_DEFAULT_HAIKU_MODEL"] = "databricks-claude-haiku-4-5"
+    settings["env"]["ANTHROPIC_CUSTOM_HEADERS"] = "x-databricks-use-coding-agent-mode: true"
+    settings["env"]["CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS"] = "1"
+
     settings_path.write_text(json.dumps(settings, indent=2))
     print(f"Claude configured: {settings_path}")
 else:

setup_codex.py

@@ -30,22 +30,39 @@
 local_bin.mkdir(parents=True, exist_ok=True)
 codex_bin = local_bin / "codex"
 
+MAX_RETRIES = 3
+RETRY_DELAY = 5  # seconds
+
 if not codex_bin.exists():
-    # Use --prefix ~/.local so npm installs directly into ~/.local/bin
     npm_prefix = str(home / ".local")
     codex_version = get_npm_version("@openai/codex")
     codex_pkg = f"@openai/codex@{codex_version}" if codex_version else "@openai/codex"
-    print(f"Installing {codex_pkg}...")
-    result = subprocess.run(
-        ["npm", "install", "-g", f"--prefix={npm_prefix}", codex_pkg],
-        capture_output=True,
-        text=True,
-        env={**os.environ, "HOME": str(home)},
-    )
-    if result.returncode == 0:
-        print(f"Codex CLI installed to {codex_bin}")
-    else:
-        print(f"Codex CLI install warning: {result.stderr}")
+
+    for attempt in range(1, MAX_RETRIES + 1):
+        print(f"Installing {codex_pkg} (attempt {attempt}/{MAX_RETRIES})...")
+        result = subprocess.run(
+            ["npm", "install", "-g", f"--prefix={npm_prefix}", codex_pkg],
+            capture_output=True,
+            text=True,
+            env={**os.environ, "HOME": str(home)},
+        )
+        if result.returncode == 0 and codex_bin.exists():
+            print(f"Codex CLI installed to {codex_bin}")
+            break
+        else:
+            stderr = result.stderr.strip()
+            print(f"Codex CLI install failed (attempt {attempt}/{MAX_RETRIES}, rc={result.returncode})")
+            if stderr:
+                print(f"  stderr: {stderr[:500]}")
+            if result.stdout.strip():
+                print(f"  stdout: {result.stdout.strip()[:500]}")
+            if attempt < MAX_RETRIES:
+                import time
+                print(f"  Retrying in {RETRY_DELAY}s...")
+                time.sleep(RETRY_DELAY)
+            else:
+                print(f"ERROR: Codex CLI installation failed after {MAX_RETRIES} attempts. "
+                      f"Run manually: npm install -g --prefix=$HOME/.local @openai/codex")
 else:
     print(f"Codex CLI already installed at {codex_bin}")
 

setup_gemini.py

@@ -12,7 +12,6 @@
 """
 import os
 import json
-import shutil
 import subprocess
 from pathlib import Path
 
@@ -33,21 +32,38 @@
 local_bin.mkdir(parents=True, exist_ok=True)
 gemini_bin = local_bin / "gemini"
 
+MAX_RETRIES = 3
+RETRY_DELAY = 5  # seconds
+
 if not gemini_bin.exists():
-    # Use --prefix ~/.local so npm installs directly into ~/.local/bin (avoids EACCES on /usr/local)
     npm_prefix = str(home / ".local")
     gemini_version = get_npm_version("@google/gemini-cli")
     gemini_pkg = f"@google/gemini-cli@{gemini_version}" if gemini_version else "@google/gemini-cli@latest"
-    print(f"Installing {gemini_pkg}...")
-    result = subprocess.run(
-        ["npm", "install", "-g", f"--prefix={npm_prefix}", gemini_pkg],
-        capture_output=True, text=True,
-        env={**os.environ, "HOME": str(home)}
-    )
-    if result.returncode == 0:
-        print(f"Gemini CLI installed to {gemini_bin}")
-    else:
-        print(f"Gemini CLI install warning: {result.stderr}")
+
+    for attempt in range(1, MAX_RETRIES + 1):
+        print(f"Installing {gemini_pkg} (attempt {attempt}/{MAX_RETRIES})...")
+        result = subprocess.run(
+            ["npm", "install", "-g", f"--prefix={npm_prefix}", gemini_pkg],
+            capture_output=True, text=True,
+            env={**os.environ, "HOME": str(home)}
+        )
+        if result.returncode == 0 and gemini_bin.exists():
+            print(f"Gemini CLI installed to {gemini_bin}")
+            break
+        else:
+            stderr = result.stderr.strip()
+            print(f"Gemini CLI install failed (attempt {attempt}/{MAX_RETRIES}, rc={result.returncode})")
+            if stderr:
+                print(f"  stderr: {stderr[:500]}")
+            if result.stdout.strip():
+                print(f"  stdout: {result.stdout.strip()[:500]}")
+            if attempt < MAX_RETRIES:
+                import time
+                print(f"  Retrying in {RETRY_DELAY}s...")
+                time.sleep(RETRY_DELAY)
+            else:
+                print(f"ERROR: Gemini CLI installation failed after {MAX_RETRIES} attempts. "
+                      f"Run manually: npm install -g --prefix=$HOME/.local @google/gemini-cli")
 else:
     print(f"Gemini CLI already installed at {gemini_bin}")
 
@@ -127,16 +143,9 @@
 settings_path.write_text(json.dumps(settings, indent=2))
 print(f"Gemini CLI settings configured: {settings_path}")
 
-# 5. Copy Claude skills into .gemini/skills for shared reference
-claude_skills_dir = home / ".claude" / "skills"
-gemini_skills_dir = gemini_dir / "skills"
-if claude_skills_dir.exists():
-    if gemini_skills_dir.exists():
-        shutil.rmtree(gemini_skills_dir)
-    shutil.copytree(claude_skills_dir, gemini_skills_dir)
-    print(f"Skills copied: {claude_skills_dir} -> {gemini_skills_dir}")
-else:
-    print(f"No Claude skills found at {claude_skills_dir}, skipping copy")
+# 5. Skills live in ~/.agents/skills/ (shared across all CLIs, copied by setup_codex.py).
+#    Do NOT copy into ~/.gemini/skills/ — Gemini discovers both paths and logs
+#    "Skill conflict detected" warnings for every duplicate.
 
 # 6. Adapt CLAUDE.md to GEMINI.md for Gemini CLI
 # Look for CLAUDE.md in common locations

setup_opencode.py

@@ -32,23 +32,40 @@
 local_bin.mkdir(parents=True, exist_ok=True)
 opencode_bin = local_bin / "opencode"
 
+MAX_RETRIES = 3
+RETRY_DELAY = 5  # seconds
+
 if not opencode_bin.exists():
-    # Use --prefix ~/.local so npm installs directly into ~/.local/bin (avoids EACCES on /usr/local)
     npm_prefix = str(home / ".local")
 
     # Resolve exact versions to avoid mutable @latest tags (supply chain hardening)
     oc_version = get_npm_version("opencode-ai")
     oc_pkg = f"opencode-ai@{oc_version}" if oc_version else "opencode-ai@latest"
-    print(f"Installing {oc_pkg}...")
-    result = subprocess.run(
-        ["npm", "install", "-g", f"--prefix={npm_prefix}", oc_pkg],
-        capture_output=True, text=True,
-        env={**os.environ, "HOME": str(home)}
-    )
-    if result.returncode == 0:
-        print(f"OpenCode CLI installed to {opencode_bin}")
-    else:
-        print(f"OpenCode install warning: {result.stderr}")
+
+    for attempt in range(1, MAX_RETRIES + 1):
+        print(f"Installing {oc_pkg} (attempt {attempt}/{MAX_RETRIES})...")
+        result = subprocess.run(
+            ["npm", "install", "-g", f"--prefix={npm_prefix}", oc_pkg],
+            capture_output=True, text=True,
+            env={**os.environ, "HOME": str(home)}
+        )
+        if result.returncode == 0 and opencode_bin.exists():
+            print(f"OpenCode CLI installed to {opencode_bin}")
+            break
+        else:
+            stderr = result.stderr.strip()
+            print(f"OpenCode install failed (attempt {attempt}/{MAX_RETRIES}, rc={result.returncode})")
+            if stderr:
+                print(f"  stderr: {stderr[:500]}")
+            if result.stdout.strip():
+                print(f"  stdout: {result.stdout.strip()[:500]}")
+            if attempt < MAX_RETRIES:
+                import time
+                print(f"  Retrying in {RETRY_DELAY}s...")
+                time.sleep(RETRY_DELAY)
+            else:
+                print(f"ERROR: OpenCode installation failed after {MAX_RETRIES} attempts. "
+                      f"Run manually: npm install -g --prefix=$HOME/.local opencode-ai")
 
     # Install @ai-sdk/openai for GPT models (Responses API support)
     sdk_version = get_npm_version("@ai-sdk/openai")
@@ -62,7 +79,7 @@
     if result.returncode == 0:
         print(f"@ai-sdk/openai@{sdk_version or 'latest'} installed (Responses API support)")
     else:
-        print(f"@ai-sdk/openai install warning: {result.stderr}")
+        print(f"@ai-sdk/openai install warning: {result.stderr[:500]}")
 else:
     print(f"OpenCode CLI already installed at {opencode_bin}")
 

tests/test_session_detach.py

@@ -105,6 +105,9 @@ class TestListSessions:
     def setup_app(self):
         app_module = _get_app()
         app_module.app_owner = "test@example.com"
+        # Clear leftover sessions from prior test classes
+        with app_module.sessions_lock:
+            app_module.sessions.clear()
         self.client = app_module.app.test_client()
         self.app_module = app_module
         yield

tests/test_session_limit.py

@@ -170,6 +170,9 @@ class TestRemovedSessionsFreeSlots:
     def test_removing_session_frees_slot(self):
         """After removing a session from the dict, a new one can be created."""
         app_module = _get_app()
+        # Clear any leaked sessions from prior tests
+        with app_module.sessions_lock:
+            app_module.sessions.clear()
         limit = app_module.MAX_CONCURRENT_SESSIONS
         sids = [f"full-{i}" for i in range(limit)]
         try: