feat: continue slow writes with standard MCP request state

.github/scripts/e2e_scope.py

@@ -24,11 +24,11 @@
 
 # Changed file -> e2e @test group(s) it exercises. Best-effort; refine as the suite evolves.
 FILE_GROUP_MAP = {
-    "libraries/mcp-native-rules-lib.groovy":    ["native_apps", "op_replay"],
+    "libraries/mcp-native-rules-lib.groovy":    ["native_apps", "mrtr"],
     "libraries/mcp-visual-rules-lib.groovy":    ["visual_rules"],
     "libraries/mcp-custom-rules-lib.groovy":    ["rule_crud", "trigger_types", "condition_types", "action_types", "complex_patterns"],
     "hubitat-mcp-rule.groovy":                  ["rule_crud", "trigger_types", "condition_types", "action_types", "complex_patterns"],
-    "libraries/mcp-diagnostics-lib.groovy":     ["diagnostics", "system_tools", "op_replay"],
+    "libraries/mcp-diagnostics-lib.groovy":     ["diagnostics", "system_tools"],
     "libraries/mcp-debug-logging-lib.groovy":   ["developer_mode", "diagnostics"],
     "libraries/mcp-system-lib.groovy":          ["system_tools", "infrastructure"],
     "libraries/mcp-self-admin-lib.groovy":      ["developer_mode", "best_practice_gating"],
@@ -42,10 +42,9 @@
     "libraries/mcp-item-backups-lib.groovy":    ["system_tools"],
     "libraries/mcp-rooms-lib.groovy":           ["infrastructure"],
     "libraries/mcp-discovery-lib.groovy":       ["infrastructure", "protocol"],
-    "libraries/mcp-app-cloner-lib.groovy":      ["native_apps", "rule_crud"],
-    "libraries/mcp-deploy-jobs-lib.groovy":     ["native_apps"],
+    "libraries/mcp-app-cloner-lib.groovy":      ["native_apps", "rule_crud", "mrtr"],
     "libraries/mcp-dashboards-lib.groovy":      ["dashboards"],
-    "hubitat-mcp-server.groovy":                ["op_replay", "protocol"],
+    "hubitat-mcp-server.groovy":                ["mrtr", "protocol"],
 }
 
 

.github/scripts/mcp_probe_hub.sh

@@ -91,12 +91,8 @@ section "hub_list_rules (RM rules via RMUtils)" "$MCP_URL" "$(gw_rpc hub_manage_
 section "hub_get_visual_rule (Visual Rules list)" "$MCP_URL" "$(gw_rpc hub_read_rules hub_get_visual_rule '{}')"
 section "hub_list_variables (hub + rule-engine)" "$MCP_URL" "$(gw_rpc hub_manage_variables hub_list_variables '{}')"
 section "hub_list_files (File Manager)" "$MCP_URL" "$(gw_rpc hub_read_files hub_list_files '{}')"
-# Targeted counts: the unfiltered listing routinely exceeds the 120KB response cap on a
-# hub with many files, which hides WHAT is accumulating. These two answer the question
-# the full listing cannot: how many op-result buffers and per-item backups are on the hub.
-# `total` in each response is the count AFTER filtering, so it is the number even when the
-# page itself is truncated.
-section "hub_list_files filter=mcp-op-result (op-token result buffers)" "$MCP_URL" "$(gw_rpc hub_read_files hub_list_files '{"filter":"mcp-op-result","includeOpResults":true}')"
+# Targeted count: the unfiltered listing routinely exceeds the 120KB response cap on a
+# hub with many files. `total` is the count after filtering, even when the page is truncated.
 section "hub_list_files filter=mcp-rm-backup (per-rule snapshots)" "$MCP_URL" "$(gw_rpc hub_read_files hub_list_files '{"filter":"mcp-rm-backup"}')"
 section "location events (lowMemory / systemStart / mode / HSM, last 24h)" "$MCP_URL" "$(tool_rpc hub_list_device_events '{"limit":50}')"
 

.github/workflows/hub-e2e.yml

@@ -411,7 +411,7 @@ jobs:
           while IFS= read -r f; do
             [ -z "$f" ] && continue
             case "$f" in
-              hubitat-mcp-server.groovy|hubitat-mcp-rule.groovy|e2e-deadman-watchdog.groovy|e2e-deadman-watchdog-v2.groovy|tests/e2e_test.py|tests/sdk_conformance_test.py|tests/sdk-conformance-requirements.txt|.github/workflows/hub-e2e.yml|.github/scripts/lease_acquire.sh|.github/scripts/lease_release.sh|.github/scripts/mcp_setup_env.sh|.github/scripts/mcp_restore_env.sh|.github/scripts/mcp_validate_package_tool.sh|.github/scripts/mcp_watchdog_lib.sh|.github/scripts/mcp_watchdog_deploy.sh|.github/scripts/mcp_arm_watchdog.sh|.github/scripts/mcp_disarm_watchdog.sh|.github/scripts/e2e_scope.py|libraries/*|bundles/*)
+              hubitat-mcp-server.groovy|hubitat-mcp-rule.groovy|e2e-deadman-watchdog.groovy|e2e-deadman-watchdog-v2.groovy|tests/e2e_test.py|tests/sdk_conformance_test.py|tests/sdk_conformance_helpers.py|tests/sdk-conformance-requirements.txt|.github/workflows/hub-e2e.yml|.github/scripts/lease_acquire.sh|.github/scripts/lease_release.sh|.github/scripts/mcp_setup_env.sh|.github/scripts/mcp_restore_env.sh|.github/scripts/mcp_validate_package_tool.sh|.github/scripts/mcp_watchdog_lib.sh|.github/scripts/mcp_watchdog_deploy.sh|.github/scripts/mcp_arm_watchdog.sh|.github/scripts/mcp_disarm_watchdog.sh|.github/scripts/e2e_scope.py|libraries/*|bundles/*)
                 relevant=true ;;
             esac
           done <<< "$files"
@@ -711,7 +711,7 @@ jobs:
           fi
 
       # Conformance leg: the official MCP Python SDK's client + validators judge the deployed
-      # PR's negotiation path (sibling: McpWireSchemaConformanceSpec; docs/testing.md).
+      # PR's pinned 2026-07-28 client path (sibling: McpWireSchemaConformanceSpec; docs/testing.md).
       #
       # Skipped under skip_install, like pkg_validate: it WRITES (one scenario flips
       # publishOutputSchemas ON and restores it in a finally), and on a skip_install dispatch it

AGENTS.md

@@ -139,7 +139,7 @@ Every MCP tool is gated. The layers, from broadest to narrowest:
 
 ### Error contracts
 
-- **Validation errors** (caller-recoverable, bad args): throw `IllegalArgumentException`. Caught by `handleToolsCall` and mapped to JSON-RPC `-32602`. Existing pattern; reaffirmed. **A validation throw MUST fire before any side effect** — the opToken machinery RELEASES a token on `-32602` (a corrected same-token re-issue executes fresh), which is only safe while nothing was committed; a tool that mutates state and then throws `IllegalArgumentException` would expose that work to a double-run.
+- **Validation errors** (caller-recoverable, bad args): throw `IllegalArgumentException`. Caught by `handleToolsCall` and mapped to JSON-RPC `-32602`. Existing pattern; reaffirmed. **A validation throw MUST fire before any side effect** — callers may safely correct and retry `-32602`, so a tool that mutates state and then throws `IllegalArgumentException` would expose that work to a double-run.
 - **Runtime errors** (operation tried and failed for non-arg reasons): return `[success: false, error: <human-readable>, note: <actionable guidance>]`. Don't throw — the AI needs a structured error.
 - **`isError: true` envelope** for tool-execution errors per MCP spec 2025-06-18. Already adopted in v0.7.7+ (see SKILL.md § Version Management for the adoption note).
 - **Specific, actionable, recovery-oriented error text.** Tell the model how to recover. Anthropic recommends steering truncation errors toward recovery strategies like *"many small and targeted searches instead of a single, broad search."*

CLAUDE.md

@@ -139,7 +139,7 @@ Every MCP tool is gated. The layers, from broadest to narrowest:
 
 ### Error contracts
 
-- **Validation errors** (caller-recoverable, bad args): throw `IllegalArgumentException`. Caught by `handleToolsCall` and mapped to JSON-RPC `-32602`. Existing pattern; reaffirmed. **A validation throw MUST fire before any side effect** — the opToken machinery RELEASES a token on `-32602` (a corrected same-token re-issue executes fresh), which is only safe while nothing was committed; a tool that mutates state and then throws `IllegalArgumentException` would expose that work to a double-run.
+- **Validation errors** (caller-recoverable, bad args): throw `IllegalArgumentException`. Caught by `handleToolsCall` and mapped to JSON-RPC `-32602`. Existing pattern; reaffirmed. **A validation throw MUST fire before any side effect** — callers may safely correct and retry `-32602`, so a tool that mutates state and then throws `IllegalArgumentException` would expose that work to a double-run.
 - **Runtime errors** (operation tried and failed for non-arg reasons): return `[success: false, error: <human-readable>, note: <actionable guidance>]`. Don't throw — the AI needs a structured error.
 - **`isError: true` envelope** for tool-execution errors per MCP spec 2025-06-18. Already adopted in v0.7.7+ (see SKILL.md § Version Management for the adoption note).
 - **Specific, actionable, recovery-oriented error text.** Tell the model how to recover. Anthropic recommends steering truncation errors toward recovery strategies like *"many small and targeted searches instead of a single, broad search."*