LearnNewsExamplesServices
Frontmatter
id10370
titleFormalize the Agent Session Sunset Protocol as a Progressive Disclosure Skill
stateClosed
labels
documentationenhancementai
assigneesneo-gemini-pro
createdAtApr 26, 2026, 2:27 PM
updatedAtApr 26, 2026, 4:34 PM
githubUrlhttps://github.com/neomjs/neo/issues/10370
authorneo-gemini-pro
commentsCount0
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
blockedBy[]
blocking[]
closedAtApr 26, 2026, 4:34 PM

Formalize the Agent Session Sunset Protocol as a Progressive Disclosure Skill

Closed v13.0.0/archive-v13-0-0-chunk-6 documentationenhancementai
neo-gemini-pro
neo-gemini-pro commented on Apr 26, 2026, 2:27 PM

Context During a recent marathon session, an agent (Claude Opus 4.7) independently executed a highly structured "Sunset Protocol" before terminating the session. This protocol included posting self-contained handover comments on relevant issues, explicitly noting considered-but-deferred handovers, documenting the mental-model state of active work phases, and summarizing marathon metrics. This emergent behavior was extremely effective but is entirely uncodified in the repository.

The Problem Sequential swarm agents suffer from "Zero-State Amnesia" (AGENTS.md §14). While we have high-level instructions about handoffs, we lack a formal, disciplined "shutdown" or "sunset" routine. Because the Sunset Protocol was hallucinated/emergent rather than codified, agents will only execute it inconsistently, leading to fragmented continuity across complex, multi-agent workflows.

The Architectural Reality The Neo.mjs Swarm requires rigorous continuity across fragmentation. We utilize a Progressive Disclosure architecture (.agent/skills/) to train agents on complex behavioral workflows without bloating the main AGENTS.md system prompt.

The Fix Formalize the emergent behavior by introducing a new session-sunset skill. This skill will provide the authoritative protocol for how an agent must terminate a complex session, ensuring the next agent has a perfect "cold-pickup" ramp. The skill should mandate the sequence:

  1. Handovers Posted: Cross-linked comments on active tickets.
  2. Handovers Considered: Explicitly documenting what not to pick up.
  3. Mental-Model State: Summary of architectural phase progress.
  4. Marathon Metrics: Turn/PR counts and derived lessons.
  5. Memory Persistence: Saving the final, rich memory node.

Acceptance Criteria

  • Create .agent/skills/session-sunset/SKILL.md (the lightweight router).
  • Create .agent/skills/session-sunset/references/session-sunset-workflow.md (the heavy payload detailing the 5-step sequence).
  • Create the required symlink: ln -sf ../../.agent/skills/session-sunset .claude/skills/session-sunset.
  • Update AGENTS.md §14 (or equivalent) to explicitly reference the session-sunset skill at session termination.

Out of Scope

  • This ticket is strictly for documenting the behavioral protocol; it does not involve modifying any MCP servers, memory core databases, or automated GitHub Actions.

Avoided Traps

  • System Prompt Bloat: We must avoid dumping the full sunset checklist into AGENTS.md. The Progressive Disclosure pattern must be used so agents only load the sunset payload when actively terminating a session.

Origin Session ID Origin Session ID: 27016011-8ae9-48bb-af87-9479dd5b0fd0

tobiu referenced in commit a945439 - "feat(ai): formalize the sunset protocol as an agent skill (#10370) (#10371) on Apr 26, 2026, 4:34 PM
tobiu closed this issue on Apr 26, 2026, 4:34 PM