Wire the agent to Command Center's secrets instead of duplicates

.github/workflows/deploy.yml

@@ -327,8 +327,14 @@ jobs:
           # single-attachment volume: rolling tries to stand up a parallel
           # machine and errors on the attachment slot. The agent group has
           # no volume, but strategy is app-wide.
+          # --ha=false because Fly provisions TWO machines by default for
+          # a newly-declared process group. For `agent` that is actively
+          # wrong: two workers drain the same run queue, and POST /start
+          # is idempotent rather than exclusive, so both would claim the
+          # same run and we would pay for the LLM call twice. The first
+          # deploy of the agent group did exactly this.
           echo "Releasing $IMAGE to sentinel-command"
-          flyctl deploy --image "$IMAGE" --strategy immediate --yes -a sentinel-command
+          flyctl deploy --image "$IMAGE" --strategy immediate --ha=false --yes -a sentinel-command
 
           echo "--- process groups after release ---"
           flyctl machines list -a sentinel-command --json \

backend/app/sentinel_agent/config.py

@@ -12,7 +12,7 @@
 the MCP server which org each tool call is on behalf of.
 """
 
-from pydantic import field_validator, model_validator
+from pydantic import AliasChoices, Field, field_validator, model_validator
 from pydantic_settings import BaseSettings
 
 
@@ -80,7 +80,25 @@ class Settings(BaseSettings):
     # The override header is still sent and is accepted as long as it
     # names the key's own org.
     opensentry_mcp_url: str = ""  # derived from opensentry_api_base if blank
-    opensentry_mcp_agent_key: str = ""
+    # Reads OPENSENTRY_MCP_AGENT_KEY first, then Command Center's own
+    # SENTINEL_AGENT_MCP_KEY. The second alias exists because the agent
+    # now runs as a process group of the sentinel-command app and shares
+    # its environment — without it the hosted agent would need a
+    # duplicate copy of a secret that is already right there, under a
+    # different name.
+    #
+    # This matters more than tidiness: the first-party deployment uses
+    # two genuinely DIFFERENT shared secrets (run-queue vs MCP), so the
+    # self-hosted fallback below would quietly hand the MCP surface the
+    # wrong key — an agent that fetches work and then fails every tool
+    # call, which is exactly the failure that fallback was written to
+    # prevent for self-hosters.
+    opensentry_mcp_agent_key: str = Field(
+        "",
+        validation_alias=AliasChoices(
+            "OPENSENTRY_MCP_AGENT_KEY", "SENTINEL_AGENT_MCP_KEY"
+        ),
+    )
 
     # ── Webhook signature behaviour ──────────────────────────────────
     # When False, /wakeup skips HMAC verification — useful for local