[Issacboy123]

Summary

Hidden subagent results and escalations can contain untrusted claims or embedded instructions, but the existing escalation-only guidance could disappear when coordination tools were filtered or a custom or extension-returned system prompt was selected. This PR gives every controller-owning root session one combined trust suffix across prompt forms and rebuild paths.

Reasoning

Applying the guidance in AgentSession after the effective system prompt is selected makes the trust boundary independent of tool availability and extension prompt replacement. Keeping the composition private avoids expanding the public prompt-building API and preserves the existing custom-message roles, persistence, delivery, and display behavior.

Changes

• Replaced tool-owned escalation guidance with one controller-bound suffix covering subagent_results and subagent_escalations.
• Appended the suffix exactly once to default, custom, and extension-returned root system prompts.
• Preserved the suffix across tool filtering, resource reloads, discovered extension resources, and harness activation.
• Added focused prompt-composition coverage for each supported prompt and rebuild path.
• Updated SDK, usage, and saved-session documentation with the delegated-report trust model and recorded the fix in the changelog.

Testing

• node ../../node_modules/vitest/dist/cli.js --run test/suite/subagent-parent-messaging.test.ts
• node ../../node_modules/vitest/dist/cli.js --run test/suite/subagent-services.test.ts test/suite/subagent-registration.test.ts test/suite/subagent-coordination.test.ts test/suite/subagent-parent-messaging.test.ts test/subagent-controller.test.ts
• npm run check

[Pepps233]

More testing should be done on this; if the parent begins validating every claim returned by the subagents, it defeats the purpose of having subagents (separation of context). This is ok to merge for now, but just keep the above in mind.