Sub-agents enter infinite loop of session lifecycle calls (mem_session_summary)

[jzambrano12]

📋 Pre-flight Checks

• I have searched existing issues and this is not a duplicate
• I understand this issue needs status:approved before a PR can be opened

📝 Bug Description

When Claude Code sub-agents (especially Explore agents) are launched by an orchestrator, they inherit the engram memory protocol from multiple injection sources — all marking SESSION CLOSE PROTOCOL as mandatory. The sub-agent interprets its task completion as a "session close" and enters an infinite loop of engram MCP calls:

session_start → mem_context → [actual work] → mem_session_summary → mem_context → mem_session_summary → session_start → mem_context → mem_session_summary → ...

The sub-agent saves the same session summary multiple times with slight variations, never terminating. With bypassPermissions enabled, all MCP calls auto-approve and the loop runs unchecked — sessions can hang for up to 1 hour until the user manually cancels.

Root cause: The engram protocol is injected into sub-agents through overlapping sources:

1. session-start.sh hook — injects into main agent context at startup
2. memory/SKILL.md — marked "ALWAYS ACTIVE", inherited by sub-agents
3. User's CLAUDE.md — may contain additional copies of the protocol

All sources say "SESSION CLOSE PROTOCOL (mandatory)" and "This is NOT optional", which sub-agents follow literally when finishing their delegated task.

🔄 Steps to Reproduce

1. Configure Claude Code with bypassPermissions enabled
2. Have the engram plugin installed and active
3. Start a new session and ask any question that triggers the orchestrator to launch an Explore sub-agent (e.g., exploring a codebase)
4. Observe the sub-agent enters an infinite loop of mem_session_start, mem_context, and mem_session_summary calls
5. The session hangs indefinitely until manually cancelled

✅ Expected Behavior

Sub-agents should complete their delegated task and return results to the parent agent without calling session lifecycle tools (mem_session_start, mem_session_end, mem_session_summary). Session lifecycle should only be managed by the top-level agent that directly interacts with the human user.

❌ Actual Behavior

Sub-agents enter an infinite loop of session lifecycle calls. Example trace from an Explore sub-agent:

plugin:engram:engram - Start Session (MCP)(id: "mp-web-session-continue", project: "mp-web")
plugin:engram:engram - Get Memory Context (MCP)(project: "mp-web", limit: 30)
plugin:engram:engram - Start Session (MCP)(id: "continue-auth-exploration-2026-03-26", project: "mp-web")
plugin:engram:engram - Get Memory Context (MCP)(project: "mp-web", limit: 30)
[... actual exploration work ...]
plugin:engram:engram - Get Memory Context (MCP)(project: "mp-web", limit: 20)
plugin:engram:engram - Save Session Summary (MCP)(project: "mp-web", content: "## Goal\n...")
plugin:engram:engram - Get Memory Context (MCP)(project: "mp-web", limit: 20)
plugin:engram:engram - Save Session Summary (MCP)(project: "mp-web", content: "## Goal\n...")  # same content, slight variation
plugin:engram:engram - Start Session (MCP)(id: "mp-web-session-continued", project: "mp-web")
plugin:engram:engram - Get Memory Context (MCP)(project: "mp-web", limit: 20)
plugin:engram:engram - Save Session Summary (MCP)(project: "mp-web", content: "## Goal\n...")  # repeats indefinitely

Workaround: Cancel the stuck operation and tell the agent "Continua" — the main agent then proceeds normally.

🖥️ Environment

• Operating System: macOS
• Engram Version: 1.10.8
• Agent / Client: Claude Code

💡 Additional Context

Proposed fix: Add a "SUB-AGENT / DELEGATED TASK SCOPE" section to the engram protocol (memory/SKILL.md and session-start.sh injected text) that explicitly tells sub-agents:

• DO NOT call mem_session_start, mem_session_end, or mem_session_summary
• DO NOT follow the SESSION CLOSE PROTOCOL
• MAY call mem_save for genuinely important discoveries (once, not repeatedly)
• MAY call mem_search or mem_context if needed for the task
• When done, return result to parent agent — that is the only "close" action

Also scope the existing headers: SESSION CLOSE PROTOCOL (mandatory) → SESSION CLOSE PROTOCOL (mandatory — TOP-LEVEL AGENT ONLY) and AFTER COMPACTION → AFTER COMPACTION (TOP-LEVEL AGENT ONLY).

This fix has been implemented and validated in gentle-ai for the engram-protocol.md asset that gets injected into CLAUDE.md. Applying the same pattern to the plugin's memory/SKILL.md and session-start.sh would fix it at the source for all users.

[jzambrano12]

@Alan-TheGentleman check this out when you have time 👋🏻