Fix SKILL.md scanner contradictions

- Qualify "Not collected" to "Not collected by OTLP telemetry" with
  explicit note that routing sends {role, content} when manifest/auto
  is active, resolving the privacy claim contradiction
- Remove MANIFEST_API_KEY and config path from requires in metadata
  since local mode doesn't need them — keeps primaryEnv for cloud users

Author: brunobuddy

.changeset/fix-skill-scanner.md

@@ -0,0 +1,5 @@
+---
+"manifest": patch
+---
+
+Fix SKILL.md scanner contradictions: qualify "Not collected" statement to clarify OTLP vs routing data, remove MANIFEST_API_KEY from required env metadata since local mode doesn't need it.

skills/manifest/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: manifest
 description: Smart LLM Router for OpenClaw. Save up to 70% by routing every request to the right model. No coding required.
-metadata: {"openclaw":{"requires":{"bins":["openclaw"],"env":["MANIFEST_API_KEY"],"config":["plugins.entries.manifest.config.apiKey"]},"primaryEnv":"MANIFEST_API_KEY","homepage":"https://github.com/mnfst/manifest"}}
+metadata: {"openclaw":{"requires":{"bins":["openclaw"]},"primaryEnv":"MANIFEST_API_KEY","homepage":"https://github.com/mnfst/manifest"}}
 ---
 
 # Manifest — LLM Router & Observability for OpenClaw
@@ -45,7 +45,7 @@ Exhaustive list of attributes sent per span:
 - `manifest.routing.reason` — routing reason (if routed)
 - Error status: agent errors truncated to 500 chars; tool errors include `event.error.message` untruncated
 
-**Not collected**: user prompts, assistant responses, tool input/output arguments, file contents, or any message body.
+**Not collected by OTLP telemetry**: user prompts, assistant responses, tool input/output arguments, file contents, or any message body. Note: when `manifest/auto` routing is active, the last 10 non-system messages (`{role, content}` only) are sent to the routing endpoint for tier scoring — see Routing Data below. To avoid this, set a fixed model instead of `manifest/auto`.
 
 ### Routing Data