Session Guard
Mandate
This skill monitors session coherence. It activates via skill-sweep on every turn — not manually invoked.
Detection Criteria
Flag the session when ANY of these hold:
- •Domain sprawl: 3+ distinct problem domains are actively in play (e.g., dotfiles config + API debugging + UI refactoring)
- •Disconnected tasks: The current task has no logical dependency on or connection to the session's first task
- •Context saturation: The conversation has been compacted/summarized AND new unrelated work is being added on top
- •Recovery drift: Error recovery or troubleshooting has shifted the session's focus permanently away from the original goal
- •Interleaved incomplete work: 2+ tasks are partially done with context for each competing for space
When NOT to Flag
- •Tasks that are logically connected steps of a single goal (e.g., "implement feature" → "write tests" → "refactor")
- •Deepening investigation within a single domain (e.g., debugging leads to architecture investigation)
- •Normal follow-up questions about completed work
Protocol
When drift is detected:
Step 1: Inform
Tell the user clearly and concisely:
Session scope notice: This session now covers [list domains]. Continuing to add unrelated work will reduce quality as context competes for space. Recommend splitting into focused sessions.
Step 2: Generate Handoff Prompt
Produce a structured prompt the user can copy-paste into a new session:
## Session Handoff — [Date] ### Context [1-2 sentence summary of what this session accomplished] ### Remaining Tasks - [ ] [Task 1 with enough context to resume] - [ ] [Task 2 with enough context to resume] ### Key Findings This Session - [Finding 1 — include file paths] - [Finding 2 — include file paths] ### Current State - Files modified: [list] - Files created: [list] - Pending: [what still needs doing] ### Reference Files - [path/to/key/file — what it contains]
Step 3: Continue or Split
After presenting the handoff prompt:
- •If the user chooses to continue, proceed — do not nag
- •If the user copies the prompt, acknowledge and focus the current session on its remaining scope
- •Never block the user from continuing; this is advisory, not enforcement
Integration with Skill-Sweep
Skill-sweep evaluates session-guard every turn. The evaluation is lightweight — it does not require research or tool calls. It is a reasoning-only check based on the conversation history visible to the agent.