AMS MCP Server Skill

Version: 3.3.0 Status: Active Updated: 2026-02-16 Module: unified-mcp-root

Use this skill to execute AMS flows safely and in the correct order.

Hard Dependency Chains

Apply these chains strictly. Do not skip prerequisites.

1. System bootstrap chain

   • unified_init -> unified_vitals -> domain tools

2. Roadmap to project chain

   • roadmap_export_to_gsd -> unified_set_project -> watcher_start

3. Audit proof chain

   • context_create or context_ensure -> audit_session_start -> work tools -> audit_verify_chain -> merkle_finalize -> merkle_root

4. Workflow execution chain

   • workflow_run dry run first for new sequences
   • If failure policy is configured: retry/backoff/rollback must be honored

5. Restoration chain

   • unified_backup before risky operations
   • Then unified_import or unified_restore
   • Then unified_vitals and audit_verify_chain for post-check

Startup Checklist

Run this at the beginning of each operational session.

1. unified_init
2. unified_vitals
3. unified_help (refresh available tools and conventions)
4. If project scoped work is expected, set project context:
   • unified_set_project or pass canonical project_id

Intent to Tool Mapping

Use this mapping to pick tools quickly.

• Explore learning paths:

  • roadmap_list, roadmap_get, roadmap_nodes, roadmap_node_get

• Validate and repair roadmap content:

  • roadmap_validate, roadmap_validate_all, roadmap_repair, roadmap_repair_all

• Create project from roadmap:

  • roadmap_export_to_gsd
  • then unified_set_project
  • then watcher_start (optional)

• Manage tasks:

  • task_create, task_create_batch, task_list, task_update, task_delete
  • task_eisenhower, task_next_actions, task_report
  • task_deps, task_link_roadmap, task_sync_to_gsd

• Analyze content:

  • analysis_search, analysis_similar, analysis_recommend, analysis_keywords, analysis_compare
  • RAG: analysis_rag_index, analysis_rag_search, analysis_rag_stats

• Manage memory and retention:

  • memory_pack, memory_get, memory_map, memory_verify, memory_bundle, memory_stats
  • memory_retention, memory_gc, rag_retention, rag_gc

• Run autonomous pipeline:

  • ams_autonomous_run, ams_autonomous_apply_changes, ams_session_list, ams_session_resume, ams_session_rollback

• Operate orchestration and architecture controls:

  • unified_orchestration, workflow_run, parallel_execute
  • unified_architecture, unified_metrics, unified_stats

• Access control:

  • acl_add_member, acl_remove_member, acl_list_members, acl_set_owner

• Notifications and git operations:

  • notify, git_checkpoint, git_auto_commit

Safety Rules

1. Always run unified_vitals before and after changes.
2. Always establish project context before project-scoped tools.
3. Never call watcher_start for a project that was not exported/initialized.
4. Use dry-run style operations first where available (repair_all, gc, import/export flows).
5. For autonomous apply mode, ensure environment policy allows it.
6. For audit-grade sessions, finalize Merkle proof before closing delivery.
7. If runtime state and tracked files diverge, stop and verify with unified_metrics and unified_help.

Common Failure Patterns

• "Project not found"

  • Cause: missing export/init or wrong project context
  • Fix: run roadmap_export_to_gsd, then unified_set_project, then retry

• "Roadmap not found"

  • Cause: invalid roadmap id
  • Fix: run roadmap_list and use exact id

• "MCP connection/runtime issue"

  • Cause: config mismatch, missing node path, server boot failure
  • Fix: verify client MCP config and run unified_vitals

• "Chain verification failed"

  • Cause: broken session/action context linkage
  • Fix: inspect with audit_get_actions, re-check context_*, then audit_verify_chain

Skill Execution Policy

When this skill is triggered:

1. Prefer deterministic tool sequences over ad-hoc calls.
2. Explain prerequisite insertion explicitly when user skips a required step.
3. Keep orchestration transparent: show what was auto-resolved and why.
4. Report concrete outputs: ids, counts, statuses, and next required step.
5. If a tool set changed, trust runtime discovery via unified_help over static counts.

Reference Loading

Load references only when needed:

1. Read references/workflow.md for end-to-end scenario flow details.
2. Read references/tools.md when the user asks for tool-specific argument examples.
3. Read references/errors.md for troubleshooting paths and recovery commands.

If references conflict with runtime behavior, runtime behavior is the source of truth.