simplify Athena runtime and centralize MCP management

platform/hermes/config.yaml

@@ -59,7 +59,7 @@ agent:
   # Keep enough deliberation for tool choice while avoiding the provider's
   # unbounded `auto` reasoning mode on ordinary turns. Users can still raise it
   # per session with /reasoning.
-  reasoning_effort: "minimal"
+  reasoning_effort: "low"
   gateway_timeout: 3600
   session_stall_timeout: 600
   tool_loop_guardrails:
@@ -77,7 +77,9 @@ agent:
       max_web_searches: 8
       max_subagents: 8
 
-# Keep ample room for long agent work. Compression starts at 82% of whichever
+# Keep ample room for long agent work. Compression starts late; deterministic
+# tool-result pruning and rolling micro-compaction keep it from getting there
+# during normal jobs.
 # router profile is selected. Keep the result compact enough that a local
 # model does not spend many minutes generating the handoff.
 compression:
@@ -89,19 +91,26 @@ compression:
   micro_compact_every_n_turns: 5
   micro_compact_defrag_threshold_tokens: 2000
   progress_notices: true
-  threshold: 0.82
+  threshold: 0.95
   target_ratio: 0.15
+  max_attempts: 1
   tail_mode: "lean"
   protect_last_n: 20
   protect_first_n: 0
-  proactive_prune_tokens: 50000
-  proactive_prune_min_result_chars: 4000
-  proactive_prune_min_reclaim_tokens: 4096
+  proactive_prune_tokens: 32000
+  proactive_prune_min_result_chars: 2000
+  proactive_prune_min_reclaim_tokens: 2048
+  context_timeout_seconds: 45
   # A failed local summarizer must not block an interactive client for ten
   # minutes. Continue without dropping messages after two minutes.
   context_total_ceiling_seconds: 120
 
 auxiliary:
+  compression:
+    provider: "main"
+    reasoning_effort: "none"
+    timeout: 120
+    max_concurrency: 1
   # Session names are cosmetic and used to create a second concurrent LLM
   # request after every first reply. On a single inference slot this blocks the
   # actual chat, so keep the original timestamp/session id instead.